Loading...
1/*
2 * Copyright (c) 2003 Patrick McHardy, <kaber@trash.net>
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version 2
7 * of the License, or (at your option) any later version.
8 *
9 * 2003-10-17 - Ported from altq
10 */
11/*
12 * Copyright (c) 1997-1999 Carnegie Mellon University. All Rights Reserved.
13 *
14 * Permission to use, copy, modify, and distribute this software and
15 * its documentation is hereby granted (including for commercial or
16 * for-profit use), provided that both the copyright notice and this
17 * permission notice appear in all copies of the software, derivative
18 * works, or modified versions, and any portions thereof.
19 *
20 * THIS SOFTWARE IS EXPERIMENTAL AND IS KNOWN TO HAVE BUGS, SOME OF
21 * WHICH MAY HAVE SERIOUS CONSEQUENCES. CARNEGIE MELLON PROVIDES THIS
22 * SOFTWARE IN ITS ``AS IS'' CONDITION, AND ANY EXPRESS OR IMPLIED
23 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
24 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
25 * DISCLAIMED. IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
27 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT
28 * OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
29 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
30 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
32 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
33 * DAMAGE.
34 *
35 * Carnegie Mellon encourages (but does not require) users of this
36 * software to return any improvements or extensions that they make,
37 * and to grant Carnegie Mellon the rights to redistribute these
38 * changes without encumbrance.
39 */
40/*
41 * H-FSC is described in Proceedings of SIGCOMM'97,
42 * "A Hierarchical Fair Service Curve Algorithm for Link-Sharing,
43 * Real-Time and Priority Service"
44 * by Ion Stoica, Hui Zhang, and T. S. Eugene Ng.
45 *
46 * Oleg Cherevko <olwi@aq.ml.com.ua> added the upperlimit for link-sharing.
47 * when a class has an upperlimit, the fit-time is computed from the
48 * upperlimit service curve. the link-sharing scheduler does not schedule
49 * a class whose fit-time exceeds the current time.
50 */
51
52#include <linux/kernel.h>
53#include <linux/module.h>
54#include <linux/types.h>
55#include <linux/errno.h>
56#include <linux/compiler.h>
57#include <linux/spinlock.h>
58#include <linux/skbuff.h>
59#include <linux/string.h>
60#include <linux/slab.h>
61#include <linux/list.h>
62#include <linux/rbtree.h>
63#include <linux/init.h>
64#include <linux/rtnetlink.h>
65#include <linux/pkt_sched.h>
66#include <net/netlink.h>
67#include <net/pkt_sched.h>
68#include <net/pkt_cls.h>
69#include <asm/div64.h>
70
71/*
72 * kernel internal service curve representation:
73 * coordinates are given by 64 bit unsigned integers.
74 * x-axis: unit is clock count.
75 * y-axis: unit is byte.
76 *
77 * The service curve parameters are converted to the internal
78 * representation. The slope values are scaled to avoid overflow.
79 * the inverse slope values as well as the y-projection of the 1st
80 * segment are kept in order to avoid 64-bit divide operations
81 * that are expensive on 32-bit architectures.
82 */
83
84struct internal_sc {
85 u64 sm1; /* scaled slope of the 1st segment */
86 u64 ism1; /* scaled inverse-slope of the 1st segment */
87 u64 dx; /* the x-projection of the 1st segment */
88 u64 dy; /* the y-projection of the 1st segment */
89 u64 sm2; /* scaled slope of the 2nd segment */
90 u64 ism2; /* scaled inverse-slope of the 2nd segment */
91};
92
93/* runtime service curve */
94struct runtime_sc {
95 u64 x; /* current starting position on x-axis */
96 u64 y; /* current starting position on y-axis */
97 u64 sm1; /* scaled slope of the 1st segment */
98 u64 ism1; /* scaled inverse-slope of the 1st segment */
99 u64 dx; /* the x-projection of the 1st segment */
100 u64 dy; /* the y-projection of the 1st segment */
101 u64 sm2; /* scaled slope of the 2nd segment */
102 u64 ism2; /* scaled inverse-slope of the 2nd segment */
103};
104
105enum hfsc_class_flags {
106 HFSC_RSC = 0x1,
107 HFSC_FSC = 0x2,
108 HFSC_USC = 0x4
109};
110
111struct hfsc_class {
112 struct Qdisc_class_common cl_common;
113 unsigned int refcnt; /* usage count */
114
115 struct gnet_stats_basic_packed bstats;
116 struct gnet_stats_queue qstats;
117 struct gnet_stats_rate_est rate_est;
118 unsigned int level; /* class level in hierarchy */
119 struct tcf_proto *filter_list; /* filter list */
120 unsigned int filter_cnt; /* filter count */
121
122 struct hfsc_sched *sched; /* scheduler data */
123 struct hfsc_class *cl_parent; /* parent class */
124 struct list_head siblings; /* sibling classes */
125 struct list_head children; /* child classes */
126 struct Qdisc *qdisc; /* leaf qdisc */
127
128 struct rb_node el_node; /* qdisc's eligible tree member */
129 struct rb_root vt_tree; /* active children sorted by cl_vt */
130 struct rb_node vt_node; /* parent's vt_tree member */
131 struct rb_root cf_tree; /* active children sorted by cl_f */
132 struct rb_node cf_node; /* parent's cf_heap member */
133 struct list_head dlist; /* drop list member */
134
135 u64 cl_total; /* total work in bytes */
136 u64 cl_cumul; /* cumulative work in bytes done by
137 real-time criteria */
138
139 u64 cl_d; /* deadline*/
140 u64 cl_e; /* eligible time */
141 u64 cl_vt; /* virtual time */
142 u64 cl_f; /* time when this class will fit for
143 link-sharing, max(myf, cfmin) */
144 u64 cl_myf; /* my fit-time (calculated from this
145 class's own upperlimit curve) */
146 u64 cl_myfadj; /* my fit-time adjustment (to cancel
147 history dependence) */
148 u64 cl_cfmin; /* earliest children's fit-time (used
149 with cl_myf to obtain cl_f) */
150 u64 cl_cvtmin; /* minimal virtual time among the
151 children fit for link-sharing
152 (monotonic within a period) */
153 u64 cl_vtadj; /* intra-period cumulative vt
154 adjustment */
155 u64 cl_vtoff; /* inter-period cumulative vt offset */
156 u64 cl_cvtmax; /* max child's vt in the last period */
157 u64 cl_cvtoff; /* cumulative cvtmax of all periods */
158 u64 cl_pcvtoff; /* parent's cvtoff at initialization
159 time */
160
161 struct internal_sc cl_rsc; /* internal real-time service curve */
162 struct internal_sc cl_fsc; /* internal fair service curve */
163 struct internal_sc cl_usc; /* internal upperlimit service curve */
164 struct runtime_sc cl_deadline; /* deadline curve */
165 struct runtime_sc cl_eligible; /* eligible curve */
166 struct runtime_sc cl_virtual; /* virtual curve */
167 struct runtime_sc cl_ulimit; /* upperlimit curve */
168
169 unsigned long cl_flags; /* which curves are valid */
170 unsigned long cl_vtperiod; /* vt period sequence number */
171 unsigned long cl_parentperiod;/* parent's vt period sequence number*/
172 unsigned long cl_nactive; /* number of active children */
173};
174
175struct hfsc_sched {
176 u16 defcls; /* default class id */
177 struct hfsc_class root; /* root class */
178 struct Qdisc_class_hash clhash; /* class hash */
179 struct rb_root eligible; /* eligible tree */
180 struct list_head droplist; /* active leaf class list (for
181 dropping) */
182 struct qdisc_watchdog watchdog; /* watchdog timer */
183};
184
185#define HT_INFINITY 0xffffffffffffffffULL /* infinite time value */
186
187
188/*
189 * eligible tree holds backlogged classes being sorted by their eligible times.
190 * there is one eligible tree per hfsc instance.
191 */
192
193static void
194eltree_insert(struct hfsc_class *cl)
195{
196 struct rb_node **p = &cl->sched->eligible.rb_node;
197 struct rb_node *parent = NULL;
198 struct hfsc_class *cl1;
199
200 while (*p != NULL) {
201 parent = *p;
202 cl1 = rb_entry(parent, struct hfsc_class, el_node);
203 if (cl->cl_e >= cl1->cl_e)
204 p = &parent->rb_right;
205 else
206 p = &parent->rb_left;
207 }
208 rb_link_node(&cl->el_node, parent, p);
209 rb_insert_color(&cl->el_node, &cl->sched->eligible);
210}
211
212static inline void
213eltree_remove(struct hfsc_class *cl)
214{
215 rb_erase(&cl->el_node, &cl->sched->eligible);
216}
217
218static inline void
219eltree_update(struct hfsc_class *cl)
220{
221 eltree_remove(cl);
222 eltree_insert(cl);
223}
224
225/* find the class with the minimum deadline among the eligible classes */
226static inline struct hfsc_class *
227eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
228{
229 struct hfsc_class *p, *cl = NULL;
230 struct rb_node *n;
231
232 for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
233 p = rb_entry(n, struct hfsc_class, el_node);
234 if (p->cl_e > cur_time)
235 break;
236 if (cl == NULL || p->cl_d < cl->cl_d)
237 cl = p;
238 }
239 return cl;
240}
241
242/* find the class with minimum eligible time among the eligible classes */
243static inline struct hfsc_class *
244eltree_get_minel(struct hfsc_sched *q)
245{
246 struct rb_node *n;
247
248 n = rb_first(&q->eligible);
249 if (n == NULL)
250 return NULL;
251 return rb_entry(n, struct hfsc_class, el_node);
252}
253
254/*
255 * vttree holds holds backlogged child classes being sorted by their virtual
256 * time. each intermediate class has one vttree.
257 */
258static void
259vttree_insert(struct hfsc_class *cl)
260{
261 struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
262 struct rb_node *parent = NULL;
263 struct hfsc_class *cl1;
264
265 while (*p != NULL) {
266 parent = *p;
267 cl1 = rb_entry(parent, struct hfsc_class, vt_node);
268 if (cl->cl_vt >= cl1->cl_vt)
269 p = &parent->rb_right;
270 else
271 p = &parent->rb_left;
272 }
273 rb_link_node(&cl->vt_node, parent, p);
274 rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
275}
276
277static inline void
278vttree_remove(struct hfsc_class *cl)
279{
280 rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
281}
282
283static inline void
284vttree_update(struct hfsc_class *cl)
285{
286 vttree_remove(cl);
287 vttree_insert(cl);
288}
289
290static inline struct hfsc_class *
291vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
292{
293 struct hfsc_class *p;
294 struct rb_node *n;
295
296 for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
297 p = rb_entry(n, struct hfsc_class, vt_node);
298 if (p->cl_f <= cur_time)
299 return p;
300 }
301 return NULL;
302}
303
304/*
305 * get the leaf class with the minimum vt in the hierarchy
306 */
307static struct hfsc_class *
308vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
309{
310 /* if root-class's cfmin is bigger than cur_time nothing to do */
311 if (cl->cl_cfmin > cur_time)
312 return NULL;
313
314 while (cl->level > 0) {
315 cl = vttree_firstfit(cl, cur_time);
316 if (cl == NULL)
317 return NULL;
318 /*
319 * update parent's cl_cvtmin.
320 */
321 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
322 cl->cl_parent->cl_cvtmin = cl->cl_vt;
323 }
324 return cl;
325}
326
327static void
328cftree_insert(struct hfsc_class *cl)
329{
330 struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
331 struct rb_node *parent = NULL;
332 struct hfsc_class *cl1;
333
334 while (*p != NULL) {
335 parent = *p;
336 cl1 = rb_entry(parent, struct hfsc_class, cf_node);
337 if (cl->cl_f >= cl1->cl_f)
338 p = &parent->rb_right;
339 else
340 p = &parent->rb_left;
341 }
342 rb_link_node(&cl->cf_node, parent, p);
343 rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
344}
345
346static inline void
347cftree_remove(struct hfsc_class *cl)
348{
349 rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
350}
351
352static inline void
353cftree_update(struct hfsc_class *cl)
354{
355 cftree_remove(cl);
356 cftree_insert(cl);
357}
358
359/*
360 * service curve support functions
361 *
362 * external service curve parameters
363 * m: bps
364 * d: us
365 * internal service curve parameters
366 * sm: (bytes/psched_us) << SM_SHIFT
367 * ism: (psched_us/byte) << ISM_SHIFT
368 * dx: psched_us
369 *
370 * The clock source resolution with ktime and PSCHED_SHIFT 10 is 1.024us.
371 *
372 * sm and ism are scaled in order to keep effective digits.
373 * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
374 * digits in decimal using the following table.
375 *
376 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
377 * ------------+-------------------------------------------------------
378 * bytes/1.024us 12.8e-3 128e-3 1280e-3 12800e-3 128000e-3
379 *
380 * 1.024us/byte 78.125 7.8125 0.78125 0.078125 0.0078125
381 *
382 * So, for PSCHED_SHIFT 10 we need: SM_SHIFT 20, ISM_SHIFT 18.
383 */
384#define SM_SHIFT (30 - PSCHED_SHIFT)
385#define ISM_SHIFT (8 + PSCHED_SHIFT)
386
387#define SM_MASK ((1ULL << SM_SHIFT) - 1)
388#define ISM_MASK ((1ULL << ISM_SHIFT) - 1)
389
390static inline u64
391seg_x2y(u64 x, u64 sm)
392{
393 u64 y;
394
395 /*
396 * compute
397 * y = x * sm >> SM_SHIFT
398 * but divide it for the upper and lower bits to avoid overflow
399 */
400 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
401 return y;
402}
403
404static inline u64
405seg_y2x(u64 y, u64 ism)
406{
407 u64 x;
408
409 if (y == 0)
410 x = 0;
411 else if (ism == HT_INFINITY)
412 x = HT_INFINITY;
413 else {
414 x = (y >> ISM_SHIFT) * ism
415 + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
416 }
417 return x;
418}
419
420/* Convert m (bps) into sm (bytes/psched us) */
421static u64
422m2sm(u32 m)
423{
424 u64 sm;
425
426 sm = ((u64)m << SM_SHIFT);
427 sm += PSCHED_TICKS_PER_SEC - 1;
428 do_div(sm, PSCHED_TICKS_PER_SEC);
429 return sm;
430}
431
432/* convert m (bps) into ism (psched us/byte) */
433static u64
434m2ism(u32 m)
435{
436 u64 ism;
437
438 if (m == 0)
439 ism = HT_INFINITY;
440 else {
441 ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
442 ism += m - 1;
443 do_div(ism, m);
444 }
445 return ism;
446}
447
448/* convert d (us) into dx (psched us) */
449static u64
450d2dx(u32 d)
451{
452 u64 dx;
453
454 dx = ((u64)d * PSCHED_TICKS_PER_SEC);
455 dx += USEC_PER_SEC - 1;
456 do_div(dx, USEC_PER_SEC);
457 return dx;
458}
459
460/* convert sm (bytes/psched us) into m (bps) */
461static u32
462sm2m(u64 sm)
463{
464 u64 m;
465
466 m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
467 return (u32)m;
468}
469
470/* convert dx (psched us) into d (us) */
471static u32
472dx2d(u64 dx)
473{
474 u64 d;
475
476 d = dx * USEC_PER_SEC;
477 do_div(d, PSCHED_TICKS_PER_SEC);
478 return (u32)d;
479}
480
481static void
482sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
483{
484 isc->sm1 = m2sm(sc->m1);
485 isc->ism1 = m2ism(sc->m1);
486 isc->dx = d2dx(sc->d);
487 isc->dy = seg_x2y(isc->dx, isc->sm1);
488 isc->sm2 = m2sm(sc->m2);
489 isc->ism2 = m2ism(sc->m2);
490}
491
492/*
493 * initialize the runtime service curve with the given internal
494 * service curve starting at (x, y).
495 */
496static void
497rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
498{
499 rtsc->x = x;
500 rtsc->y = y;
501 rtsc->sm1 = isc->sm1;
502 rtsc->ism1 = isc->ism1;
503 rtsc->dx = isc->dx;
504 rtsc->dy = isc->dy;
505 rtsc->sm2 = isc->sm2;
506 rtsc->ism2 = isc->ism2;
507}
508
509/*
510 * calculate the y-projection of the runtime service curve by the
511 * given x-projection value
512 */
513static u64
514rtsc_y2x(struct runtime_sc *rtsc, u64 y)
515{
516 u64 x;
517
518 if (y < rtsc->y)
519 x = rtsc->x;
520 else if (y <= rtsc->y + rtsc->dy) {
521 /* x belongs to the 1st segment */
522 if (rtsc->dy == 0)
523 x = rtsc->x + rtsc->dx;
524 else
525 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
526 } else {
527 /* x belongs to the 2nd segment */
528 x = rtsc->x + rtsc->dx
529 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
530 }
531 return x;
532}
533
534static u64
535rtsc_x2y(struct runtime_sc *rtsc, u64 x)
536{
537 u64 y;
538
539 if (x <= rtsc->x)
540 y = rtsc->y;
541 else if (x <= rtsc->x + rtsc->dx)
542 /* y belongs to the 1st segment */
543 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
544 else
545 /* y belongs to the 2nd segment */
546 y = rtsc->y + rtsc->dy
547 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
548 return y;
549}
550
551/*
552 * update the runtime service curve by taking the minimum of the current
553 * runtime service curve and the service curve starting at (x, y).
554 */
555static void
556rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
557{
558 u64 y1, y2, dx, dy;
559 u32 dsm;
560
561 if (isc->sm1 <= isc->sm2) {
562 /* service curve is convex */
563 y1 = rtsc_x2y(rtsc, x);
564 if (y1 < y)
565 /* the current rtsc is smaller */
566 return;
567 rtsc->x = x;
568 rtsc->y = y;
569 return;
570 }
571
572 /*
573 * service curve is concave
574 * compute the two y values of the current rtsc
575 * y1: at x
576 * y2: at (x + dx)
577 */
578 y1 = rtsc_x2y(rtsc, x);
579 if (y1 <= y) {
580 /* rtsc is below isc, no change to rtsc */
581 return;
582 }
583
584 y2 = rtsc_x2y(rtsc, x + isc->dx);
585 if (y2 >= y + isc->dy) {
586 /* rtsc is above isc, replace rtsc by isc */
587 rtsc->x = x;
588 rtsc->y = y;
589 rtsc->dx = isc->dx;
590 rtsc->dy = isc->dy;
591 return;
592 }
593
594 /*
595 * the two curves intersect
596 * compute the offsets (dx, dy) using the reverse
597 * function of seg_x2y()
598 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
599 */
600 dx = (y1 - y) << SM_SHIFT;
601 dsm = isc->sm1 - isc->sm2;
602 do_div(dx, dsm);
603 /*
604 * check if (x, y1) belongs to the 1st segment of rtsc.
605 * if so, add the offset.
606 */
607 if (rtsc->x + rtsc->dx > x)
608 dx += rtsc->x + rtsc->dx - x;
609 dy = seg_x2y(dx, isc->sm1);
610
611 rtsc->x = x;
612 rtsc->y = y;
613 rtsc->dx = dx;
614 rtsc->dy = dy;
615}
616
617static void
618init_ed(struct hfsc_class *cl, unsigned int next_len)
619{
620 u64 cur_time = psched_get_time();
621
622 /* update the deadline curve */
623 rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
624
625 /*
626 * update the eligible curve.
627 * for concave, it is equal to the deadline curve.
628 * for convex, it is a linear curve with slope m2.
629 */
630 cl->cl_eligible = cl->cl_deadline;
631 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
632 cl->cl_eligible.dx = 0;
633 cl->cl_eligible.dy = 0;
634 }
635
636 /* compute e and d */
637 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
638 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
639
640 eltree_insert(cl);
641}
642
643static void
644update_ed(struct hfsc_class *cl, unsigned int next_len)
645{
646 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
647 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
648
649 eltree_update(cl);
650}
651
652static inline void
653update_d(struct hfsc_class *cl, unsigned int next_len)
654{
655 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
656}
657
658static inline void
659update_cfmin(struct hfsc_class *cl)
660{
661 struct rb_node *n = rb_first(&cl->cf_tree);
662 struct hfsc_class *p;
663
664 if (n == NULL) {
665 cl->cl_cfmin = 0;
666 return;
667 }
668 p = rb_entry(n, struct hfsc_class, cf_node);
669 cl->cl_cfmin = p->cl_f;
670}
671
672static void
673init_vf(struct hfsc_class *cl, unsigned int len)
674{
675 struct hfsc_class *max_cl;
676 struct rb_node *n;
677 u64 vt, f, cur_time;
678 int go_active;
679
680 cur_time = 0;
681 go_active = 1;
682 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
683 if (go_active && cl->cl_nactive++ == 0)
684 go_active = 1;
685 else
686 go_active = 0;
687
688 if (go_active) {
689 n = rb_last(&cl->cl_parent->vt_tree);
690 if (n != NULL) {
691 max_cl = rb_entry(n, struct hfsc_class, vt_node);
692 /*
693 * set vt to the average of the min and max
694 * classes. if the parent's period didn't
695 * change, don't decrease vt of the class.
696 */
697 vt = max_cl->cl_vt;
698 if (cl->cl_parent->cl_cvtmin != 0)
699 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
700
701 if (cl->cl_parent->cl_vtperiod !=
702 cl->cl_parentperiod || vt > cl->cl_vt)
703 cl->cl_vt = vt;
704 } else {
705 /*
706 * first child for a new parent backlog period.
707 * add parent's cvtmax to cvtoff to make a new
708 * vt (vtoff + vt) larger than the vt in the
709 * last period for all children.
710 */
711 vt = cl->cl_parent->cl_cvtmax;
712 cl->cl_parent->cl_cvtoff += vt;
713 cl->cl_parent->cl_cvtmax = 0;
714 cl->cl_parent->cl_cvtmin = 0;
715 cl->cl_vt = 0;
716 }
717
718 cl->cl_vtoff = cl->cl_parent->cl_cvtoff -
719 cl->cl_pcvtoff;
720
721 /* update the virtual curve */
722 vt = cl->cl_vt + cl->cl_vtoff;
723 rtsc_min(&cl->cl_virtual, &cl->cl_fsc, vt,
724 cl->cl_total);
725 if (cl->cl_virtual.x == vt) {
726 cl->cl_virtual.x -= cl->cl_vtoff;
727 cl->cl_vtoff = 0;
728 }
729 cl->cl_vtadj = 0;
730
731 cl->cl_vtperiod++; /* increment vt period */
732 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
733 if (cl->cl_parent->cl_nactive == 0)
734 cl->cl_parentperiod++;
735 cl->cl_f = 0;
736
737 vttree_insert(cl);
738 cftree_insert(cl);
739
740 if (cl->cl_flags & HFSC_USC) {
741 /* class has upper limit curve */
742 if (cur_time == 0)
743 cur_time = psched_get_time();
744
745 /* update the ulimit curve */
746 rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
747 cl->cl_total);
748 /* compute myf */
749 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
750 cl->cl_total);
751 cl->cl_myfadj = 0;
752 }
753 }
754
755 f = max(cl->cl_myf, cl->cl_cfmin);
756 if (f != cl->cl_f) {
757 cl->cl_f = f;
758 cftree_update(cl);
759 }
760 update_cfmin(cl->cl_parent);
761 }
762}
763
764static void
765update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
766{
767 u64 f; /* , myf_bound, delta; */
768 int go_passive = 0;
769
770 if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
771 go_passive = 1;
772
773 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
774 cl->cl_total += len;
775
776 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
777 continue;
778
779 if (go_passive && --cl->cl_nactive == 0)
780 go_passive = 1;
781 else
782 go_passive = 0;
783
784 if (go_passive) {
785 /* no more active child, going passive */
786
787 /* update cvtmax of the parent class */
788 if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
789 cl->cl_parent->cl_cvtmax = cl->cl_vt;
790
791 /* remove this class from the vt tree */
792 vttree_remove(cl);
793
794 cftree_remove(cl);
795 update_cfmin(cl->cl_parent);
796
797 continue;
798 }
799
800 /*
801 * update vt and f
802 */
803 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
804 - cl->cl_vtoff + cl->cl_vtadj;
805
806 /*
807 * if vt of the class is smaller than cvtmin,
808 * the class was skipped in the past due to non-fit.
809 * if so, we need to adjust vtadj.
810 */
811 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
812 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
813 cl->cl_vt = cl->cl_parent->cl_cvtmin;
814 }
815
816 /* update the vt tree */
817 vttree_update(cl);
818
819 if (cl->cl_flags & HFSC_USC) {
820 cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
821 cl->cl_total);
822#if 0
823 /*
824 * This code causes classes to stay way under their
825 * limit when multiple classes are used at gigabit
826 * speed. needs investigation. -kaber
827 */
828 /*
829 * if myf lags behind by more than one clock tick
830 * from the current time, adjust myfadj to prevent
831 * a rate-limited class from going greedy.
832 * in a steady state under rate-limiting, myf
833 * fluctuates within one clock tick.
834 */
835 myf_bound = cur_time - PSCHED_JIFFIE2US(1);
836 if (cl->cl_myf < myf_bound) {
837 delta = cur_time - cl->cl_myf;
838 cl->cl_myfadj += delta;
839 cl->cl_myf += delta;
840 }
841#endif
842 }
843
844 f = max(cl->cl_myf, cl->cl_cfmin);
845 if (f != cl->cl_f) {
846 cl->cl_f = f;
847 cftree_update(cl);
848 update_cfmin(cl->cl_parent);
849 }
850 }
851}
852
853static void
854set_active(struct hfsc_class *cl, unsigned int len)
855{
856 if (cl->cl_flags & HFSC_RSC)
857 init_ed(cl, len);
858 if (cl->cl_flags & HFSC_FSC)
859 init_vf(cl, len);
860
861 list_add_tail(&cl->dlist, &cl->sched->droplist);
862}
863
864static void
865set_passive(struct hfsc_class *cl)
866{
867 if (cl->cl_flags & HFSC_RSC)
868 eltree_remove(cl);
869
870 list_del(&cl->dlist);
871
872 /*
873 * vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
874 * needs to be called explicitly to remove a class from vttree.
875 */
876}
877
878static unsigned int
879qdisc_peek_len(struct Qdisc *sch)
880{
881 struct sk_buff *skb;
882 unsigned int len;
883
884 skb = sch->ops->peek(sch);
885 if (skb == NULL) {
886 qdisc_warn_nonwc("qdisc_peek_len", sch);
887 return 0;
888 }
889 len = qdisc_pkt_len(skb);
890
891 return len;
892}
893
894static void
895hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl)
896{
897 unsigned int len = cl->qdisc->q.qlen;
898
899 qdisc_reset(cl->qdisc);
900 qdisc_tree_decrease_qlen(cl->qdisc, len);
901}
902
903static void
904hfsc_adjust_levels(struct hfsc_class *cl)
905{
906 struct hfsc_class *p;
907 unsigned int level;
908
909 do {
910 level = 0;
911 list_for_each_entry(p, &cl->children, siblings) {
912 if (p->level >= level)
913 level = p->level + 1;
914 }
915 cl->level = level;
916 } while ((cl = cl->cl_parent) != NULL);
917}
918
919static inline struct hfsc_class *
920hfsc_find_class(u32 classid, struct Qdisc *sch)
921{
922 struct hfsc_sched *q = qdisc_priv(sch);
923 struct Qdisc_class_common *clc;
924
925 clc = qdisc_class_find(&q->clhash, classid);
926 if (clc == NULL)
927 return NULL;
928 return container_of(clc, struct hfsc_class, cl_common);
929}
930
931static void
932hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
933 u64 cur_time)
934{
935 sc2isc(rsc, &cl->cl_rsc);
936 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
937 cl->cl_eligible = cl->cl_deadline;
938 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
939 cl->cl_eligible.dx = 0;
940 cl->cl_eligible.dy = 0;
941 }
942 cl->cl_flags |= HFSC_RSC;
943}
944
945static void
946hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
947{
948 sc2isc(fsc, &cl->cl_fsc);
949 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
950 cl->cl_flags |= HFSC_FSC;
951}
952
953static void
954hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
955 u64 cur_time)
956{
957 sc2isc(usc, &cl->cl_usc);
958 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
959 cl->cl_flags |= HFSC_USC;
960}
961
962static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
963 [TCA_HFSC_RSC] = { .len = sizeof(struct tc_service_curve) },
964 [TCA_HFSC_FSC] = { .len = sizeof(struct tc_service_curve) },
965 [TCA_HFSC_USC] = { .len = sizeof(struct tc_service_curve) },
966};
967
968static int
969hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
970 struct nlattr **tca, unsigned long *arg)
971{
972 struct hfsc_sched *q = qdisc_priv(sch);
973 struct hfsc_class *cl = (struct hfsc_class *)*arg;
974 struct hfsc_class *parent = NULL;
975 struct nlattr *opt = tca[TCA_OPTIONS];
976 struct nlattr *tb[TCA_HFSC_MAX + 1];
977 struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
978 u64 cur_time;
979 int err;
980
981 if (opt == NULL)
982 return -EINVAL;
983
984 err = nla_parse_nested(tb, TCA_HFSC_MAX, opt, hfsc_policy);
985 if (err < 0)
986 return err;
987
988 if (tb[TCA_HFSC_RSC]) {
989 rsc = nla_data(tb[TCA_HFSC_RSC]);
990 if (rsc->m1 == 0 && rsc->m2 == 0)
991 rsc = NULL;
992 }
993
994 if (tb[TCA_HFSC_FSC]) {
995 fsc = nla_data(tb[TCA_HFSC_FSC]);
996 if (fsc->m1 == 0 && fsc->m2 == 0)
997 fsc = NULL;
998 }
999
1000 if (tb[TCA_HFSC_USC]) {
1001 usc = nla_data(tb[TCA_HFSC_USC]);
1002 if (usc->m1 == 0 && usc->m2 == 0)
1003 usc = NULL;
1004 }
1005
1006 if (cl != NULL) {
1007 if (parentid) {
1008 if (cl->cl_parent &&
1009 cl->cl_parent->cl_common.classid != parentid)
1010 return -EINVAL;
1011 if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
1012 return -EINVAL;
1013 }
1014 cur_time = psched_get_time();
1015
1016 if (tca[TCA_RATE]) {
1017 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
1018 qdisc_root_sleeping_lock(sch),
1019 tca[TCA_RATE]);
1020 if (err)
1021 return err;
1022 }
1023
1024 sch_tree_lock(sch);
1025 if (rsc != NULL)
1026 hfsc_change_rsc(cl, rsc, cur_time);
1027 if (fsc != NULL)
1028 hfsc_change_fsc(cl, fsc);
1029 if (usc != NULL)
1030 hfsc_change_usc(cl, usc, cur_time);
1031
1032 if (cl->qdisc->q.qlen != 0) {
1033 if (cl->cl_flags & HFSC_RSC)
1034 update_ed(cl, qdisc_peek_len(cl->qdisc));
1035 if (cl->cl_flags & HFSC_FSC)
1036 update_vf(cl, 0, cur_time);
1037 }
1038 sch_tree_unlock(sch);
1039
1040 return 0;
1041 }
1042
1043 if (parentid == TC_H_ROOT)
1044 return -EEXIST;
1045
1046 parent = &q->root;
1047 if (parentid) {
1048 parent = hfsc_find_class(parentid, sch);
1049 if (parent == NULL)
1050 return -ENOENT;
1051 }
1052
1053 if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1054 return -EINVAL;
1055 if (hfsc_find_class(classid, sch))
1056 return -EEXIST;
1057
1058 if (rsc == NULL && fsc == NULL)
1059 return -EINVAL;
1060
1061 cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1062 if (cl == NULL)
1063 return -ENOBUFS;
1064
1065 if (tca[TCA_RATE]) {
1066 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
1067 qdisc_root_sleeping_lock(sch),
1068 tca[TCA_RATE]);
1069 if (err) {
1070 kfree(cl);
1071 return err;
1072 }
1073 }
1074
1075 if (rsc != NULL)
1076 hfsc_change_rsc(cl, rsc, 0);
1077 if (fsc != NULL)
1078 hfsc_change_fsc(cl, fsc);
1079 if (usc != NULL)
1080 hfsc_change_usc(cl, usc, 0);
1081
1082 cl->cl_common.classid = classid;
1083 cl->refcnt = 1;
1084 cl->sched = q;
1085 cl->cl_parent = parent;
1086 cl->qdisc = qdisc_create_dflt(sch->dev_queue,
1087 &pfifo_qdisc_ops, classid);
1088 if (cl->qdisc == NULL)
1089 cl->qdisc = &noop_qdisc;
1090 INIT_LIST_HEAD(&cl->children);
1091 cl->vt_tree = RB_ROOT;
1092 cl->cf_tree = RB_ROOT;
1093
1094 sch_tree_lock(sch);
1095 qdisc_class_hash_insert(&q->clhash, &cl->cl_common);
1096 list_add_tail(&cl->siblings, &parent->children);
1097 if (parent->level == 0)
1098 hfsc_purge_queue(sch, parent);
1099 hfsc_adjust_levels(parent);
1100 cl->cl_pcvtoff = parent->cl_cvtoff;
1101 sch_tree_unlock(sch);
1102
1103 qdisc_class_hash_grow(sch, &q->clhash);
1104
1105 *arg = (unsigned long)cl;
1106 return 0;
1107}
1108
1109static void
1110hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1111{
1112 struct hfsc_sched *q = qdisc_priv(sch);
1113
1114 tcf_destroy_chain(&cl->filter_list);
1115 qdisc_destroy(cl->qdisc);
1116 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1117 if (cl != &q->root)
1118 kfree(cl);
1119}
1120
1121static int
1122hfsc_delete_class(struct Qdisc *sch, unsigned long arg)
1123{
1124 struct hfsc_sched *q = qdisc_priv(sch);
1125 struct hfsc_class *cl = (struct hfsc_class *)arg;
1126
1127 if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1128 return -EBUSY;
1129
1130 sch_tree_lock(sch);
1131
1132 list_del(&cl->siblings);
1133 hfsc_adjust_levels(cl->cl_parent);
1134
1135 hfsc_purge_queue(sch, cl);
1136 qdisc_class_hash_remove(&q->clhash, &cl->cl_common);
1137
1138 BUG_ON(--cl->refcnt == 0);
1139 /*
1140 * This shouldn't happen: we "hold" one cops->get() when called
1141 * from tc_ctl_tclass; the destroy method is done from cops->put().
1142 */
1143
1144 sch_tree_unlock(sch);
1145 return 0;
1146}
1147
1148static struct hfsc_class *
1149hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1150{
1151 struct hfsc_sched *q = qdisc_priv(sch);
1152 struct hfsc_class *head, *cl;
1153 struct tcf_result res;
1154 struct tcf_proto *tcf;
1155 int result;
1156
1157 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1158 (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1159 if (cl->level == 0)
1160 return cl;
1161
1162 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
1163 head = &q->root;
1164 tcf = q->root.filter_list;
1165 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
1166#ifdef CONFIG_NET_CLS_ACT
1167 switch (result) {
1168 case TC_ACT_QUEUED:
1169 case TC_ACT_STOLEN:
1170 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
1171 case TC_ACT_SHOT:
1172 return NULL;
1173 }
1174#endif
1175 cl = (struct hfsc_class *)res.class;
1176 if (!cl) {
1177 cl = hfsc_find_class(res.classid, sch);
1178 if (!cl)
1179 break; /* filter selected invalid classid */
1180 if (cl->level >= head->level)
1181 break; /* filter may only point downwards */
1182 }
1183
1184 if (cl->level == 0)
1185 return cl; /* hit leaf class */
1186
1187 /* apply inner filter chain */
1188 tcf = cl->filter_list;
1189 head = cl;
1190 }
1191
1192 /* classification failed, try default class */
1193 cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1194 if (cl == NULL || cl->level > 0)
1195 return NULL;
1196
1197 return cl;
1198}
1199
1200static int
1201hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1202 struct Qdisc **old)
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);
1211 if (new == NULL)
1212 new = &noop_qdisc;
1213 }
1214
1215 sch_tree_lock(sch);
1216 hfsc_purge_queue(sch, cl);
1217 *old = cl->qdisc;
1218 cl->qdisc = new;
1219 sch_tree_unlock(sch);
1220 return 0;
1221}
1222
1223static struct Qdisc *
1224hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1225{
1226 struct hfsc_class *cl = (struct hfsc_class *)arg;
1227
1228 if (cl->level == 0)
1229 return cl->qdisc;
1230
1231 return NULL;
1232}
1233
1234static void
1235hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1236{
1237 struct hfsc_class *cl = (struct hfsc_class *)arg;
1238
1239 if (cl->qdisc->q.qlen == 0) {
1240 update_vf(cl, 0, 0);
1241 set_passive(cl);
1242 }
1243}
1244
1245static unsigned long
1246hfsc_get_class(struct Qdisc *sch, u32 classid)
1247{
1248 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1249
1250 if (cl != NULL)
1251 cl->refcnt++;
1252
1253 return (unsigned long)cl;
1254}
1255
1256static void
1257hfsc_put_class(struct Qdisc *sch, unsigned long arg)
1258{
1259 struct hfsc_class *cl = (struct hfsc_class *)arg;
1260
1261 if (--cl->refcnt == 0)
1262 hfsc_destroy_class(sch, cl);
1263}
1264
1265static unsigned long
1266hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1267{
1268 struct hfsc_class *p = (struct hfsc_class *)parent;
1269 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1270
1271 if (cl != NULL) {
1272 if (p != NULL && p->level <= cl->level)
1273 return 0;
1274 cl->filter_cnt++;
1275 }
1276
1277 return (unsigned long)cl;
1278}
1279
1280static void
1281hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1282{
1283 struct hfsc_class *cl = (struct hfsc_class *)arg;
1284
1285 cl->filter_cnt--;
1286}
1287
1288static struct tcf_proto **
1289hfsc_tcf_chain(struct Qdisc *sch, unsigned long arg)
1290{
1291 struct hfsc_sched *q = qdisc_priv(sch);
1292 struct hfsc_class *cl = (struct hfsc_class *)arg;
1293
1294 if (cl == NULL)
1295 cl = &q->root;
1296
1297 return &cl->filter_list;
1298}
1299
1300static int
1301hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1302{
1303 struct tc_service_curve tsc;
1304
1305 tsc.m1 = sm2m(sc->sm1);
1306 tsc.d = dx2d(sc->dx);
1307 tsc.m2 = sm2m(sc->sm2);
1308 if (nla_put(skb, attr, sizeof(tsc), &tsc))
1309 goto nla_put_failure;
1310
1311 return skb->len;
1312
1313 nla_put_failure:
1314 return -1;
1315}
1316
1317static int
1318hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1319{
1320 if ((cl->cl_flags & HFSC_RSC) &&
1321 (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1322 goto nla_put_failure;
1323
1324 if ((cl->cl_flags & HFSC_FSC) &&
1325 (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1326 goto nla_put_failure;
1327
1328 if ((cl->cl_flags & HFSC_USC) &&
1329 (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1330 goto nla_put_failure;
1331
1332 return skb->len;
1333
1334 nla_put_failure:
1335 return -1;
1336}
1337
1338static int
1339hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1340 struct tcmsg *tcm)
1341{
1342 struct hfsc_class *cl = (struct hfsc_class *)arg;
1343 struct nlattr *nest;
1344
1345 tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid :
1346 TC_H_ROOT;
1347 tcm->tcm_handle = cl->cl_common.classid;
1348 if (cl->level == 0)
1349 tcm->tcm_info = cl->qdisc->handle;
1350
1351 nest = nla_nest_start(skb, TCA_OPTIONS);
1352 if (nest == NULL)
1353 goto nla_put_failure;
1354 if (hfsc_dump_curves(skb, cl) < 0)
1355 goto nla_put_failure;
1356 nla_nest_end(skb, nest);
1357 return skb->len;
1358
1359 nla_put_failure:
1360 nla_nest_cancel(skb, nest);
1361 return -EMSGSIZE;
1362}
1363
1364static int
1365hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1366 struct gnet_dump *d)
1367{
1368 struct hfsc_class *cl = (struct hfsc_class *)arg;
1369 struct tc_hfsc_stats xstats;
1370
1371 cl->qstats.qlen = cl->qdisc->q.qlen;
1372 cl->qstats.backlog = cl->qdisc->qstats.backlog;
1373 xstats.level = cl->level;
1374 xstats.period = cl->cl_vtperiod;
1375 xstats.work = cl->cl_total;
1376 xstats.rtwork = cl->cl_cumul;
1377
1378 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1379 gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
1380 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1381 return -1;
1382
1383 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1384}
1385
1386
1387
1388static void
1389hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1390{
1391 struct hfsc_sched *q = qdisc_priv(sch);
1392 struct hlist_node *n;
1393 struct hfsc_class *cl;
1394 unsigned int i;
1395
1396 if (arg->stop)
1397 return;
1398
1399 for (i = 0; i < q->clhash.hashsize; i++) {
1400 hlist_for_each_entry(cl, n, &q->clhash.hash[i],
1401 cl_common.hnode) {
1402 if (arg->count < arg->skip) {
1403 arg->count++;
1404 continue;
1405 }
1406 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1407 arg->stop = 1;
1408 return;
1409 }
1410 arg->count++;
1411 }
1412 }
1413}
1414
1415static void
1416hfsc_schedule_watchdog(struct Qdisc *sch)
1417{
1418 struct hfsc_sched *q = qdisc_priv(sch);
1419 struct hfsc_class *cl;
1420 u64 next_time = 0;
1421
1422 cl = eltree_get_minel(q);
1423 if (cl)
1424 next_time = cl->cl_e;
1425 if (q->root.cl_cfmin != 0) {
1426 if (next_time == 0 || next_time > q->root.cl_cfmin)
1427 next_time = q->root.cl_cfmin;
1428 }
1429 WARN_ON(next_time == 0);
1430 qdisc_watchdog_schedule(&q->watchdog, next_time);
1431}
1432
1433static int
1434hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1435{
1436 struct hfsc_sched *q = qdisc_priv(sch);
1437 struct tc_hfsc_qopt *qopt;
1438 int err;
1439
1440 if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1441 return -EINVAL;
1442 qopt = nla_data(opt);
1443
1444 q->defcls = qopt->defcls;
1445 err = qdisc_class_hash_init(&q->clhash);
1446 if (err < 0)
1447 return err;
1448 q->eligible = RB_ROOT;
1449 INIT_LIST_HEAD(&q->droplist);
1450
1451 q->root.cl_common.classid = sch->handle;
1452 q->root.refcnt = 1;
1453 q->root.sched = q;
1454 q->root.qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1455 sch->handle);
1456 if (q->root.qdisc == NULL)
1457 q->root.qdisc = &noop_qdisc;
1458 INIT_LIST_HEAD(&q->root.children);
1459 q->root.vt_tree = RB_ROOT;
1460 q->root.cf_tree = RB_ROOT;
1461
1462 qdisc_class_hash_insert(&q->clhash, &q->root.cl_common);
1463 qdisc_class_hash_grow(sch, &q->clhash);
1464
1465 qdisc_watchdog_init(&q->watchdog, sch);
1466
1467 return 0;
1468}
1469
1470static int
1471hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt)
1472{
1473 struct hfsc_sched *q = qdisc_priv(sch);
1474 struct tc_hfsc_qopt *qopt;
1475
1476 if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1477 return -EINVAL;
1478 qopt = nla_data(opt);
1479
1480 sch_tree_lock(sch);
1481 q->defcls = qopt->defcls;
1482 sch_tree_unlock(sch);
1483
1484 return 0;
1485}
1486
1487static void
1488hfsc_reset_class(struct hfsc_class *cl)
1489{
1490 cl->cl_total = 0;
1491 cl->cl_cumul = 0;
1492 cl->cl_d = 0;
1493 cl->cl_e = 0;
1494 cl->cl_vt = 0;
1495 cl->cl_vtadj = 0;
1496 cl->cl_vtoff = 0;
1497 cl->cl_cvtmin = 0;
1498 cl->cl_cvtmax = 0;
1499 cl->cl_cvtoff = 0;
1500 cl->cl_pcvtoff = 0;
1501 cl->cl_vtperiod = 0;
1502 cl->cl_parentperiod = 0;
1503 cl->cl_f = 0;
1504 cl->cl_myf = 0;
1505 cl->cl_myfadj = 0;
1506 cl->cl_cfmin = 0;
1507 cl->cl_nactive = 0;
1508
1509 cl->vt_tree = RB_ROOT;
1510 cl->cf_tree = RB_ROOT;
1511 qdisc_reset(cl->qdisc);
1512
1513 if (cl->cl_flags & HFSC_RSC)
1514 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1515 if (cl->cl_flags & HFSC_FSC)
1516 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1517 if (cl->cl_flags & HFSC_USC)
1518 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1519}
1520
1521static void
1522hfsc_reset_qdisc(struct Qdisc *sch)
1523{
1524 struct hfsc_sched *q = qdisc_priv(sch);
1525 struct hfsc_class *cl;
1526 struct hlist_node *n;
1527 unsigned int i;
1528
1529 for (i = 0; i < q->clhash.hashsize; i++) {
1530 hlist_for_each_entry(cl, n, &q->clhash.hash[i], cl_common.hnode)
1531 hfsc_reset_class(cl);
1532 }
1533 q->eligible = RB_ROOT;
1534 INIT_LIST_HEAD(&q->droplist);
1535 qdisc_watchdog_cancel(&q->watchdog);
1536 sch->q.qlen = 0;
1537}
1538
1539static void
1540hfsc_destroy_qdisc(struct Qdisc *sch)
1541{
1542 struct hfsc_sched *q = qdisc_priv(sch);
1543 struct hlist_node *n, *next;
1544 struct hfsc_class *cl;
1545 unsigned int i;
1546
1547 for (i = 0; i < q->clhash.hashsize; i++) {
1548 hlist_for_each_entry(cl, n, &q->clhash.hash[i], cl_common.hnode)
1549 tcf_destroy_chain(&cl->filter_list);
1550 }
1551 for (i = 0; i < q->clhash.hashsize; i++) {
1552 hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[i],
1553 cl_common.hnode)
1554 hfsc_destroy_class(sch, cl);
1555 }
1556 qdisc_class_hash_destroy(&q->clhash);
1557 qdisc_watchdog_cancel(&q->watchdog);
1558}
1559
1560static int
1561hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1562{
1563 struct hfsc_sched *q = qdisc_priv(sch);
1564 unsigned char *b = skb_tail_pointer(skb);
1565 struct tc_hfsc_qopt qopt;
1566 struct hfsc_class *cl;
1567 struct hlist_node *n;
1568 unsigned int i;
1569
1570 sch->qstats.backlog = 0;
1571 for (i = 0; i < q->clhash.hashsize; i++) {
1572 hlist_for_each_entry(cl, n, &q->clhash.hash[i], cl_common.hnode)
1573 sch->qstats.backlog += cl->qdisc->qstats.backlog;
1574 }
1575
1576 qopt.defcls = q->defcls;
1577 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1578 goto nla_put_failure;
1579 return skb->len;
1580
1581 nla_put_failure:
1582 nlmsg_trim(skb, b);
1583 return -1;
1584}
1585
1586static int
1587hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1588{
1589 struct hfsc_class *cl;
1590 int uninitialized_var(err);
1591
1592 cl = hfsc_classify(skb, sch, &err);
1593 if (cl == NULL) {
1594 if (err & __NET_XMIT_BYPASS)
1595 sch->qstats.drops++;
1596 kfree_skb(skb);
1597 return err;
1598 }
1599
1600 err = qdisc_enqueue(skb, cl->qdisc);
1601 if (unlikely(err != NET_XMIT_SUCCESS)) {
1602 if (net_xmit_drop_count(err)) {
1603 cl->qstats.drops++;
1604 sch->qstats.drops++;
1605 }
1606 return err;
1607 }
1608
1609 if (cl->qdisc->q.qlen == 1)
1610 set_active(cl, qdisc_pkt_len(skb));
1611
1612 sch->q.qlen++;
1613
1614 return NET_XMIT_SUCCESS;
1615}
1616
1617static struct sk_buff *
1618hfsc_dequeue(struct Qdisc *sch)
1619{
1620 struct hfsc_sched *q = qdisc_priv(sch);
1621 struct hfsc_class *cl;
1622 struct sk_buff *skb;
1623 u64 cur_time;
1624 unsigned int next_len;
1625 int realtime = 0;
1626
1627 if (sch->q.qlen == 0)
1628 return NULL;
1629
1630 cur_time = psched_get_time();
1631
1632 /*
1633 * if there are eligible classes, use real-time criteria.
1634 * find the class with the minimum deadline among
1635 * the eligible classes.
1636 */
1637 cl = eltree_get_mindl(q, cur_time);
1638 if (cl) {
1639 realtime = 1;
1640 } else {
1641 /*
1642 * use link-sharing criteria
1643 * get the class with the minimum vt in the hierarchy
1644 */
1645 cl = vttree_get_minvt(&q->root, cur_time);
1646 if (cl == NULL) {
1647 sch->qstats.overlimits++;
1648 hfsc_schedule_watchdog(sch);
1649 return NULL;
1650 }
1651 }
1652
1653 skb = qdisc_dequeue_peeked(cl->qdisc);
1654 if (skb == NULL) {
1655 qdisc_warn_nonwc("HFSC", cl->qdisc);
1656 return NULL;
1657 }
1658
1659 bstats_update(&cl->bstats, skb);
1660 update_vf(cl, qdisc_pkt_len(skb), cur_time);
1661 if (realtime)
1662 cl->cl_cumul += qdisc_pkt_len(skb);
1663
1664 if (cl->qdisc->q.qlen != 0) {
1665 if (cl->cl_flags & HFSC_RSC) {
1666 /* update ed */
1667 next_len = qdisc_peek_len(cl->qdisc);
1668 if (realtime)
1669 update_ed(cl, next_len);
1670 else
1671 update_d(cl, next_len);
1672 }
1673 } else {
1674 /* the class becomes passive */
1675 set_passive(cl);
1676 }
1677
1678 qdisc_unthrottled(sch);
1679 qdisc_bstats_update(sch, skb);
1680 sch->q.qlen--;
1681
1682 return skb;
1683}
1684
1685static unsigned int
1686hfsc_drop(struct Qdisc *sch)
1687{
1688 struct hfsc_sched *q = qdisc_priv(sch);
1689 struct hfsc_class *cl;
1690 unsigned int len;
1691
1692 list_for_each_entry(cl, &q->droplist, dlist) {
1693 if (cl->qdisc->ops->drop != NULL &&
1694 (len = cl->qdisc->ops->drop(cl->qdisc)) > 0) {
1695 if (cl->qdisc->q.qlen == 0) {
1696 update_vf(cl, 0, 0);
1697 set_passive(cl);
1698 } else {
1699 list_move_tail(&cl->dlist, &q->droplist);
1700 }
1701 cl->qstats.drops++;
1702 sch->qstats.drops++;
1703 sch->q.qlen--;
1704 return len;
1705 }
1706 }
1707 return 0;
1708}
1709
1710static const struct Qdisc_class_ops hfsc_class_ops = {
1711 .change = hfsc_change_class,
1712 .delete = hfsc_delete_class,
1713 .graft = hfsc_graft_class,
1714 .leaf = hfsc_class_leaf,
1715 .qlen_notify = hfsc_qlen_notify,
1716 .get = hfsc_get_class,
1717 .put = hfsc_put_class,
1718 .bind_tcf = hfsc_bind_tcf,
1719 .unbind_tcf = hfsc_unbind_tcf,
1720 .tcf_chain = hfsc_tcf_chain,
1721 .dump = hfsc_dump_class,
1722 .dump_stats = hfsc_dump_class_stats,
1723 .walk = hfsc_walk
1724};
1725
1726static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1727 .id = "hfsc",
1728 .init = hfsc_init_qdisc,
1729 .change = hfsc_change_qdisc,
1730 .reset = hfsc_reset_qdisc,
1731 .destroy = hfsc_destroy_qdisc,
1732 .dump = hfsc_dump_qdisc,
1733 .enqueue = hfsc_enqueue,
1734 .dequeue = hfsc_dequeue,
1735 .peek = qdisc_peek_dequeued,
1736 .drop = hfsc_drop,
1737 .cl_ops = &hfsc_class_ops,
1738 .priv_size = sizeof(struct hfsc_sched),
1739 .owner = THIS_MODULE
1740};
1741
1742static int __init
1743hfsc_init(void)
1744{
1745 return register_qdisc(&hfsc_qdisc_ops);
1746}
1747
1748static void __exit
1749hfsc_cleanup(void)
1750{
1751 unregister_qdisc(&hfsc_qdisc_ops);
1752}
1753
1754MODULE_LICENSE("GPL");
1755module_init(hfsc_init);
1756module_exit(hfsc_cleanup);
1/*
2 * Copyright (c) 2003 Patrick McHardy, <kaber@trash.net>
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version 2
7 * of the License, or (at your option) any later version.
8 *
9 * 2003-10-17 - Ported from altq
10 */
11/*
12 * Copyright (c) 1997-1999 Carnegie Mellon University. All Rights Reserved.
13 *
14 * Permission to use, copy, modify, and distribute this software and
15 * its documentation is hereby granted (including for commercial or
16 * for-profit use), provided that both the copyright notice and this
17 * permission notice appear in all copies of the software, derivative
18 * works, or modified versions, and any portions thereof.
19 *
20 * THIS SOFTWARE IS EXPERIMENTAL AND IS KNOWN TO HAVE BUGS, SOME OF
21 * WHICH MAY HAVE SERIOUS CONSEQUENCES. CARNEGIE MELLON PROVIDES THIS
22 * SOFTWARE IN ITS ``AS IS'' CONDITION, AND ANY EXPRESS OR IMPLIED
23 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
24 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
25 * DISCLAIMED. IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
27 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT
28 * OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
29 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
30 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
32 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
33 * DAMAGE.
34 *
35 * Carnegie Mellon encourages (but does not require) users of this
36 * software to return any improvements or extensions that they make,
37 * and to grant Carnegie Mellon the rights to redistribute these
38 * changes without encumbrance.
39 */
40/*
41 * H-FSC is described in Proceedings of SIGCOMM'97,
42 * "A Hierarchical Fair Service Curve Algorithm for Link-Sharing,
43 * Real-Time and Priority Service"
44 * by Ion Stoica, Hui Zhang, and T. S. Eugene Ng.
45 *
46 * Oleg Cherevko <olwi@aq.ml.com.ua> added the upperlimit for link-sharing.
47 * when a class has an upperlimit, the fit-time is computed from the
48 * upperlimit service curve. the link-sharing scheduler does not schedule
49 * a class whose fit-time exceeds the current time.
50 */
51
52#include <linux/kernel.h>
53#include <linux/module.h>
54#include <linux/types.h>
55#include <linux/errno.h>
56#include <linux/compiler.h>
57#include <linux/spinlock.h>
58#include <linux/skbuff.h>
59#include <linux/string.h>
60#include <linux/slab.h>
61#include <linux/list.h>
62#include <linux/rbtree.h>
63#include <linux/init.h>
64#include <linux/rtnetlink.h>
65#include <linux/pkt_sched.h>
66#include <net/netlink.h>
67#include <net/pkt_sched.h>
68#include <net/pkt_cls.h>
69#include <asm/div64.h>
70
71/*
72 * kernel internal service curve representation:
73 * coordinates are given by 64 bit unsigned integers.
74 * x-axis: unit is clock count.
75 * y-axis: unit is byte.
76 *
77 * The service curve parameters are converted to the internal
78 * representation. The slope values are scaled to avoid overflow.
79 * the inverse slope values as well as the y-projection of the 1st
80 * segment are kept in order to avoid 64-bit divide operations
81 * that are expensive on 32-bit architectures.
82 */
83
84struct internal_sc {
85 u64 sm1; /* scaled slope of the 1st segment */
86 u64 ism1; /* scaled inverse-slope of the 1st segment */
87 u64 dx; /* the x-projection of the 1st segment */
88 u64 dy; /* the y-projection of the 1st segment */
89 u64 sm2; /* scaled slope of the 2nd segment */
90 u64 ism2; /* scaled inverse-slope of the 2nd segment */
91};
92
93/* runtime service curve */
94struct runtime_sc {
95 u64 x; /* current starting position on x-axis */
96 u64 y; /* current starting position on y-axis */
97 u64 sm1; /* scaled slope of the 1st segment */
98 u64 ism1; /* scaled inverse-slope of the 1st segment */
99 u64 dx; /* the x-projection of the 1st segment */
100 u64 dy; /* the y-projection of the 1st segment */
101 u64 sm2; /* scaled slope of the 2nd segment */
102 u64 ism2; /* scaled inverse-slope of the 2nd segment */
103};
104
105enum hfsc_class_flags {
106 HFSC_RSC = 0x1,
107 HFSC_FSC = 0x2,
108 HFSC_USC = 0x4
109};
110
111struct hfsc_class {
112 struct Qdisc_class_common cl_common;
113
114 struct gnet_stats_basic_packed 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 filter_cnt; /* filter count */
120 unsigned int level; /* class level in hierarchy */
121
122 struct hfsc_sched *sched; /* scheduler data */
123 struct hfsc_class *cl_parent; /* parent class */
124 struct list_head siblings; /* sibling classes */
125 struct list_head children; /* child classes */
126 struct Qdisc *qdisc; /* leaf qdisc */
127
128 struct rb_node el_node; /* qdisc's eligible tree member */
129 struct rb_root vt_tree; /* active children sorted by cl_vt */
130 struct rb_node vt_node; /* parent's vt_tree member */
131 struct rb_root cf_tree; /* active children sorted by cl_f */
132 struct rb_node cf_node; /* parent's cf_heap member */
133
134 u64 cl_total; /* total work in bytes */
135 u64 cl_cumul; /* cumulative work in bytes done by
136 real-time criteria */
137
138 u64 cl_d; /* deadline*/
139 u64 cl_e; /* eligible time */
140 u64 cl_vt; /* virtual time */
141 u64 cl_f; /* time when this class will fit for
142 link-sharing, max(myf, cfmin) */
143 u64 cl_myf; /* my fit-time (calculated from this
144 class's own upperlimit curve) */
145 u64 cl_cfmin; /* earliest children's fit-time (used
146 with cl_myf to obtain cl_f) */
147 u64 cl_cvtmin; /* minimal virtual time among the
148 children fit for link-sharing
149 (monotonic within a period) */
150 u64 cl_vtadj; /* intra-period cumulative vt
151 adjustment */
152 u64 cl_cvtoff; /* largest virtual time seen among
153 the children */
154
155 struct internal_sc cl_rsc; /* internal real-time service curve */
156 struct internal_sc cl_fsc; /* internal fair service curve */
157 struct internal_sc cl_usc; /* internal upperlimit service curve */
158 struct runtime_sc cl_deadline; /* deadline curve */
159 struct runtime_sc cl_eligible; /* eligible curve */
160 struct runtime_sc cl_virtual; /* virtual curve */
161 struct runtime_sc cl_ulimit; /* upperlimit curve */
162
163 u8 cl_flags; /* which curves are valid */
164 u32 cl_vtperiod; /* vt period sequence number */
165 u32 cl_parentperiod;/* parent's vt period sequence number*/
166 u32 cl_nactive; /* number of active children */
167};
168
169struct hfsc_sched {
170 u16 defcls; /* default class id */
171 struct hfsc_class root; /* root class */
172 struct Qdisc_class_hash clhash; /* class hash */
173 struct rb_root eligible; /* eligible tree */
174 struct qdisc_watchdog watchdog; /* watchdog timer */
175};
176
177#define HT_INFINITY 0xffffffffffffffffULL /* infinite time value */
178
179
180/*
181 * eligible tree holds backlogged classes being sorted by their eligible times.
182 * there is one eligible tree per hfsc instance.
183 */
184
185static void
186eltree_insert(struct hfsc_class *cl)
187{
188 struct rb_node **p = &cl->sched->eligible.rb_node;
189 struct rb_node *parent = NULL;
190 struct hfsc_class *cl1;
191
192 while (*p != NULL) {
193 parent = *p;
194 cl1 = rb_entry(parent, struct hfsc_class, el_node);
195 if (cl->cl_e >= cl1->cl_e)
196 p = &parent->rb_right;
197 else
198 p = &parent->rb_left;
199 }
200 rb_link_node(&cl->el_node, parent, p);
201 rb_insert_color(&cl->el_node, &cl->sched->eligible);
202}
203
204static inline void
205eltree_remove(struct hfsc_class *cl)
206{
207 rb_erase(&cl->el_node, &cl->sched->eligible);
208}
209
210static inline void
211eltree_update(struct hfsc_class *cl)
212{
213 eltree_remove(cl);
214 eltree_insert(cl);
215}
216
217/* find the class with the minimum deadline among the eligible classes */
218static inline struct hfsc_class *
219eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
220{
221 struct hfsc_class *p, *cl = NULL;
222 struct rb_node *n;
223
224 for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
225 p = rb_entry(n, struct hfsc_class, el_node);
226 if (p->cl_e > cur_time)
227 break;
228 if (cl == NULL || p->cl_d < cl->cl_d)
229 cl = p;
230 }
231 return cl;
232}
233
234/* find the class with minimum eligible time among the eligible classes */
235static inline struct hfsc_class *
236eltree_get_minel(struct hfsc_sched *q)
237{
238 struct rb_node *n;
239
240 n = rb_first(&q->eligible);
241 if (n == NULL)
242 return NULL;
243 return rb_entry(n, struct hfsc_class, el_node);
244}
245
246/*
247 * vttree holds holds backlogged child classes being sorted by their virtual
248 * time. each intermediate class has one vttree.
249 */
250static void
251vttree_insert(struct hfsc_class *cl)
252{
253 struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
254 struct rb_node *parent = NULL;
255 struct hfsc_class *cl1;
256
257 while (*p != NULL) {
258 parent = *p;
259 cl1 = rb_entry(parent, struct hfsc_class, vt_node);
260 if (cl->cl_vt >= cl1->cl_vt)
261 p = &parent->rb_right;
262 else
263 p = &parent->rb_left;
264 }
265 rb_link_node(&cl->vt_node, parent, p);
266 rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
267}
268
269static inline void
270vttree_remove(struct hfsc_class *cl)
271{
272 rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
273}
274
275static inline void
276vttree_update(struct hfsc_class *cl)
277{
278 vttree_remove(cl);
279 vttree_insert(cl);
280}
281
282static inline struct hfsc_class *
283vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
284{
285 struct hfsc_class *p;
286 struct rb_node *n;
287
288 for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
289 p = rb_entry(n, struct hfsc_class, vt_node);
290 if (p->cl_f <= cur_time)
291 return p;
292 }
293 return NULL;
294}
295
296/*
297 * get the leaf class with the minimum vt in the hierarchy
298 */
299static struct hfsc_class *
300vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
301{
302 /* if root-class's cfmin is bigger than cur_time nothing to do */
303 if (cl->cl_cfmin > cur_time)
304 return NULL;
305
306 while (cl->level > 0) {
307 cl = vttree_firstfit(cl, cur_time);
308 if (cl == NULL)
309 return NULL;
310 /*
311 * update parent's cl_cvtmin.
312 */
313 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
314 cl->cl_parent->cl_cvtmin = cl->cl_vt;
315 }
316 return cl;
317}
318
319static void
320cftree_insert(struct hfsc_class *cl)
321{
322 struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
323 struct rb_node *parent = NULL;
324 struct hfsc_class *cl1;
325
326 while (*p != NULL) {
327 parent = *p;
328 cl1 = rb_entry(parent, struct hfsc_class, cf_node);
329 if (cl->cl_f >= cl1->cl_f)
330 p = &parent->rb_right;
331 else
332 p = &parent->rb_left;
333 }
334 rb_link_node(&cl->cf_node, parent, p);
335 rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
336}
337
338static inline void
339cftree_remove(struct hfsc_class *cl)
340{
341 rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
342}
343
344static inline void
345cftree_update(struct hfsc_class *cl)
346{
347 cftree_remove(cl);
348 cftree_insert(cl);
349}
350
351/*
352 * service curve support functions
353 *
354 * external service curve parameters
355 * m: bps
356 * d: us
357 * internal service curve parameters
358 * sm: (bytes/psched_us) << SM_SHIFT
359 * ism: (psched_us/byte) << ISM_SHIFT
360 * dx: psched_us
361 *
362 * The clock source resolution with ktime and PSCHED_SHIFT 10 is 1.024us.
363 *
364 * sm and ism are scaled in order to keep effective digits.
365 * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
366 * digits in decimal using the following table.
367 *
368 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
369 * ------------+-------------------------------------------------------
370 * bytes/1.024us 12.8e-3 128e-3 1280e-3 12800e-3 128000e-3
371 *
372 * 1.024us/byte 78.125 7.8125 0.78125 0.078125 0.0078125
373 *
374 * So, for PSCHED_SHIFT 10 we need: SM_SHIFT 20, ISM_SHIFT 18.
375 */
376#define SM_SHIFT (30 - PSCHED_SHIFT)
377#define ISM_SHIFT (8 + PSCHED_SHIFT)
378
379#define SM_MASK ((1ULL << SM_SHIFT) - 1)
380#define ISM_MASK ((1ULL << ISM_SHIFT) - 1)
381
382static inline u64
383seg_x2y(u64 x, u64 sm)
384{
385 u64 y;
386
387 /*
388 * compute
389 * y = x * sm >> SM_SHIFT
390 * but divide it for the upper and lower bits to avoid overflow
391 */
392 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
393 return y;
394}
395
396static inline u64
397seg_y2x(u64 y, u64 ism)
398{
399 u64 x;
400
401 if (y == 0)
402 x = 0;
403 else if (ism == HT_INFINITY)
404 x = HT_INFINITY;
405 else {
406 x = (y >> ISM_SHIFT) * ism
407 + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
408 }
409 return x;
410}
411
412/* Convert m (bps) into sm (bytes/psched us) */
413static u64
414m2sm(u32 m)
415{
416 u64 sm;
417
418 sm = ((u64)m << SM_SHIFT);
419 sm += PSCHED_TICKS_PER_SEC - 1;
420 do_div(sm, PSCHED_TICKS_PER_SEC);
421 return sm;
422}
423
424/* convert m (bps) into ism (psched us/byte) */
425static u64
426m2ism(u32 m)
427{
428 u64 ism;
429
430 if (m == 0)
431 ism = HT_INFINITY;
432 else {
433 ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
434 ism += m - 1;
435 do_div(ism, m);
436 }
437 return ism;
438}
439
440/* convert d (us) into dx (psched us) */
441static u64
442d2dx(u32 d)
443{
444 u64 dx;
445
446 dx = ((u64)d * PSCHED_TICKS_PER_SEC);
447 dx += USEC_PER_SEC - 1;
448 do_div(dx, USEC_PER_SEC);
449 return dx;
450}
451
452/* convert sm (bytes/psched us) into m (bps) */
453static u32
454sm2m(u64 sm)
455{
456 u64 m;
457
458 m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
459 return (u32)m;
460}
461
462/* convert dx (psched us) into d (us) */
463static u32
464dx2d(u64 dx)
465{
466 u64 d;
467
468 d = dx * USEC_PER_SEC;
469 do_div(d, PSCHED_TICKS_PER_SEC);
470 return (u32)d;
471}
472
473static void
474sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
475{
476 isc->sm1 = m2sm(sc->m1);
477 isc->ism1 = m2ism(sc->m1);
478 isc->dx = d2dx(sc->d);
479 isc->dy = seg_x2y(isc->dx, isc->sm1);
480 isc->sm2 = m2sm(sc->m2);
481 isc->ism2 = m2ism(sc->m2);
482}
483
484/*
485 * initialize the runtime service curve with the given internal
486 * service curve starting at (x, y).
487 */
488static void
489rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
490{
491 rtsc->x = x;
492 rtsc->y = y;
493 rtsc->sm1 = isc->sm1;
494 rtsc->ism1 = isc->ism1;
495 rtsc->dx = isc->dx;
496 rtsc->dy = isc->dy;
497 rtsc->sm2 = isc->sm2;
498 rtsc->ism2 = isc->ism2;
499}
500
501/*
502 * calculate the y-projection of the runtime service curve by the
503 * given x-projection value
504 */
505static u64
506rtsc_y2x(struct runtime_sc *rtsc, u64 y)
507{
508 u64 x;
509
510 if (y < rtsc->y)
511 x = rtsc->x;
512 else if (y <= rtsc->y + rtsc->dy) {
513 /* x belongs to the 1st segment */
514 if (rtsc->dy == 0)
515 x = rtsc->x + rtsc->dx;
516 else
517 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
518 } else {
519 /* x belongs to the 2nd segment */
520 x = rtsc->x + rtsc->dx
521 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
522 }
523 return x;
524}
525
526static u64
527rtsc_x2y(struct runtime_sc *rtsc, u64 x)
528{
529 u64 y;
530
531 if (x <= rtsc->x)
532 y = rtsc->y;
533 else if (x <= rtsc->x + rtsc->dx)
534 /* y belongs to the 1st segment */
535 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
536 else
537 /* y belongs to the 2nd segment */
538 y = rtsc->y + rtsc->dy
539 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
540 return y;
541}
542
543/*
544 * update the runtime service curve by taking the minimum of the current
545 * runtime service curve and the service curve starting at (x, y).
546 */
547static void
548rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
549{
550 u64 y1, y2, dx, dy;
551 u32 dsm;
552
553 if (isc->sm1 <= isc->sm2) {
554 /* service curve is convex */
555 y1 = rtsc_x2y(rtsc, x);
556 if (y1 < y)
557 /* the current rtsc is smaller */
558 return;
559 rtsc->x = x;
560 rtsc->y = y;
561 return;
562 }
563
564 /*
565 * service curve is concave
566 * compute the two y values of the current rtsc
567 * y1: at x
568 * y2: at (x + dx)
569 */
570 y1 = rtsc_x2y(rtsc, x);
571 if (y1 <= y) {
572 /* rtsc is below isc, no change to rtsc */
573 return;
574 }
575
576 y2 = rtsc_x2y(rtsc, x + isc->dx);
577 if (y2 >= y + isc->dy) {
578 /* rtsc is above isc, replace rtsc by isc */
579 rtsc->x = x;
580 rtsc->y = y;
581 rtsc->dx = isc->dx;
582 rtsc->dy = isc->dy;
583 return;
584 }
585
586 /*
587 * the two curves intersect
588 * compute the offsets (dx, dy) using the reverse
589 * function of seg_x2y()
590 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
591 */
592 dx = (y1 - y) << SM_SHIFT;
593 dsm = isc->sm1 - isc->sm2;
594 do_div(dx, dsm);
595 /*
596 * check if (x, y1) belongs to the 1st segment of rtsc.
597 * if so, add the offset.
598 */
599 if (rtsc->x + rtsc->dx > x)
600 dx += rtsc->x + rtsc->dx - x;
601 dy = seg_x2y(dx, isc->sm1);
602
603 rtsc->x = x;
604 rtsc->y = y;
605 rtsc->dx = dx;
606 rtsc->dy = dy;
607}
608
609static void
610init_ed(struct hfsc_class *cl, unsigned int next_len)
611{
612 u64 cur_time = psched_get_time();
613
614 /* update the deadline curve */
615 rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
616
617 /*
618 * update the eligible curve.
619 * for concave, it is equal to the deadline curve.
620 * for convex, it is a linear curve with slope m2.
621 */
622 cl->cl_eligible = cl->cl_deadline;
623 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
624 cl->cl_eligible.dx = 0;
625 cl->cl_eligible.dy = 0;
626 }
627
628 /* compute e and d */
629 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
630 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
631
632 eltree_insert(cl);
633}
634
635static void
636update_ed(struct hfsc_class *cl, unsigned int next_len)
637{
638 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
639 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
640
641 eltree_update(cl);
642}
643
644static inline void
645update_d(struct hfsc_class *cl, unsigned int next_len)
646{
647 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
648}
649
650static inline void
651update_cfmin(struct hfsc_class *cl)
652{
653 struct rb_node *n = rb_first(&cl->cf_tree);
654 struct hfsc_class *p;
655
656 if (n == NULL) {
657 cl->cl_cfmin = 0;
658 return;
659 }
660 p = rb_entry(n, struct hfsc_class, cf_node);
661 cl->cl_cfmin = p->cl_f;
662}
663
664static void
665init_vf(struct hfsc_class *cl, unsigned int len)
666{
667 struct hfsc_class *max_cl;
668 struct rb_node *n;
669 u64 vt, f, cur_time;
670 int go_active;
671
672 cur_time = 0;
673 go_active = 1;
674 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
675 if (go_active && cl->cl_nactive++ == 0)
676 go_active = 1;
677 else
678 go_active = 0;
679
680 if (go_active) {
681 n = rb_last(&cl->cl_parent->vt_tree);
682 if (n != NULL) {
683 max_cl = rb_entry(n, struct hfsc_class, vt_node);
684 /*
685 * set vt to the average of the min and max
686 * classes. if the parent's period didn't
687 * change, don't decrease vt of the class.
688 */
689 vt = max_cl->cl_vt;
690 if (cl->cl_parent->cl_cvtmin != 0)
691 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
692
693 if (cl->cl_parent->cl_vtperiod !=
694 cl->cl_parentperiod || vt > cl->cl_vt)
695 cl->cl_vt = vt;
696 } else {
697 /*
698 * first child for a new parent backlog period.
699 * initialize cl_vt to the highest value seen
700 * among the siblings. this is analogous to
701 * what cur_time would provide in realtime case.
702 */
703 cl->cl_vt = cl->cl_parent->cl_cvtoff;
704 cl->cl_parent->cl_cvtmin = 0;
705 }
706
707 /* update the virtual curve */
708 rtsc_min(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
709 cl->cl_vtadj = 0;
710
711 cl->cl_vtperiod++; /* increment vt period */
712 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
713 if (cl->cl_parent->cl_nactive == 0)
714 cl->cl_parentperiod++;
715 cl->cl_f = 0;
716
717 vttree_insert(cl);
718 cftree_insert(cl);
719
720 if (cl->cl_flags & HFSC_USC) {
721 /* class has upper limit curve */
722 if (cur_time == 0)
723 cur_time = psched_get_time();
724
725 /* update the ulimit curve */
726 rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
727 cl->cl_total);
728 /* compute myf */
729 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
730 cl->cl_total);
731 }
732 }
733
734 f = max(cl->cl_myf, cl->cl_cfmin);
735 if (f != cl->cl_f) {
736 cl->cl_f = f;
737 cftree_update(cl);
738 }
739 update_cfmin(cl->cl_parent);
740 }
741}
742
743static void
744update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
745{
746 u64 f; /* , myf_bound, delta; */
747 int go_passive = 0;
748
749 if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
750 go_passive = 1;
751
752 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
753 cl->cl_total += len;
754
755 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
756 continue;
757
758 if (go_passive && --cl->cl_nactive == 0)
759 go_passive = 1;
760 else
761 go_passive = 0;
762
763 /* update vt */
764 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total) + cl->cl_vtadj;
765
766 /*
767 * if vt of the class is smaller than cvtmin,
768 * the class was skipped in the past due to non-fit.
769 * if so, we need to adjust vtadj.
770 */
771 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
772 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
773 cl->cl_vt = cl->cl_parent->cl_cvtmin;
774 }
775
776 if (go_passive) {
777 /* no more active child, going passive */
778
779 /* update cvtoff of the parent class */
780 if (cl->cl_vt > cl->cl_parent->cl_cvtoff)
781 cl->cl_parent->cl_cvtoff = cl->cl_vt;
782
783 /* remove this class from the vt tree */
784 vttree_remove(cl);
785
786 cftree_remove(cl);
787 update_cfmin(cl->cl_parent);
788
789 continue;
790 }
791
792 /* update the vt tree */
793 vttree_update(cl);
794
795 /* update f */
796 if (cl->cl_flags & HFSC_USC) {
797 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit, cl->cl_total);
798#if 0
799 cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
800 cl->cl_total);
801 /*
802 * This code causes classes to stay way under their
803 * limit when multiple classes are used at gigabit
804 * speed. needs investigation. -kaber
805 */
806 /*
807 * if myf lags behind by more than one clock tick
808 * from the current time, adjust myfadj to prevent
809 * a rate-limited class from going greedy.
810 * in a steady state under rate-limiting, myf
811 * fluctuates within one clock tick.
812 */
813 myf_bound = cur_time - PSCHED_JIFFIE2US(1);
814 if (cl->cl_myf < myf_bound) {
815 delta = cur_time - cl->cl_myf;
816 cl->cl_myfadj += delta;
817 cl->cl_myf += delta;
818 }
819#endif
820 }
821
822 f = max(cl->cl_myf, cl->cl_cfmin);
823 if (f != cl->cl_f) {
824 cl->cl_f = f;
825 cftree_update(cl);
826 update_cfmin(cl->cl_parent);
827 }
828 }
829}
830
831static unsigned int
832qdisc_peek_len(struct Qdisc *sch)
833{
834 struct sk_buff *skb;
835 unsigned int len;
836
837 skb = sch->ops->peek(sch);
838 if (unlikely(skb == NULL)) {
839 qdisc_warn_nonwc("qdisc_peek_len", sch);
840 return 0;
841 }
842 len = qdisc_pkt_len(skb);
843
844 return len;
845}
846
847static void
848hfsc_adjust_levels(struct hfsc_class *cl)
849{
850 struct hfsc_class *p;
851 unsigned int level;
852
853 do {
854 level = 0;
855 list_for_each_entry(p, &cl->children, siblings) {
856 if (p->level >= level)
857 level = p->level + 1;
858 }
859 cl->level = level;
860 } while ((cl = cl->cl_parent) != NULL);
861}
862
863static inline struct hfsc_class *
864hfsc_find_class(u32 classid, struct Qdisc *sch)
865{
866 struct hfsc_sched *q = qdisc_priv(sch);
867 struct Qdisc_class_common *clc;
868
869 clc = qdisc_class_find(&q->clhash, classid);
870 if (clc == NULL)
871 return NULL;
872 return container_of(clc, struct hfsc_class, cl_common);
873}
874
875static void
876hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
877 u64 cur_time)
878{
879 sc2isc(rsc, &cl->cl_rsc);
880 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
881 cl->cl_eligible = cl->cl_deadline;
882 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
883 cl->cl_eligible.dx = 0;
884 cl->cl_eligible.dy = 0;
885 }
886 cl->cl_flags |= HFSC_RSC;
887}
888
889static void
890hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
891{
892 sc2isc(fsc, &cl->cl_fsc);
893 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
894 cl->cl_flags |= HFSC_FSC;
895}
896
897static void
898hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
899 u64 cur_time)
900{
901 sc2isc(usc, &cl->cl_usc);
902 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
903 cl->cl_flags |= HFSC_USC;
904}
905
906static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
907 [TCA_HFSC_RSC] = { .len = sizeof(struct tc_service_curve) },
908 [TCA_HFSC_FSC] = { .len = sizeof(struct tc_service_curve) },
909 [TCA_HFSC_USC] = { .len = sizeof(struct tc_service_curve) },
910};
911
912static int
913hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
914 struct nlattr **tca, unsigned long *arg,
915 struct netlink_ext_ack *extack)
916{
917 struct hfsc_sched *q = qdisc_priv(sch);
918 struct hfsc_class *cl = (struct hfsc_class *)*arg;
919 struct hfsc_class *parent = NULL;
920 struct nlattr *opt = tca[TCA_OPTIONS];
921 struct nlattr *tb[TCA_HFSC_MAX + 1];
922 struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
923 u64 cur_time;
924 int err;
925
926 if (opt == NULL)
927 return -EINVAL;
928
929 err = nla_parse_nested_deprecated(tb, TCA_HFSC_MAX, opt, hfsc_policy,
930 NULL);
931 if (err < 0)
932 return err;
933
934 if (tb[TCA_HFSC_RSC]) {
935 rsc = nla_data(tb[TCA_HFSC_RSC]);
936 if (rsc->m1 == 0 && rsc->m2 == 0)
937 rsc = NULL;
938 }
939
940 if (tb[TCA_HFSC_FSC]) {
941 fsc = nla_data(tb[TCA_HFSC_FSC]);
942 if (fsc->m1 == 0 && fsc->m2 == 0)
943 fsc = NULL;
944 }
945
946 if (tb[TCA_HFSC_USC]) {
947 usc = nla_data(tb[TCA_HFSC_USC]);
948 if (usc->m1 == 0 && usc->m2 == 0)
949 usc = NULL;
950 }
951
952 if (cl != NULL) {
953 int old_flags;
954
955 if (parentid) {
956 if (cl->cl_parent &&
957 cl->cl_parent->cl_common.classid != parentid)
958 return -EINVAL;
959 if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
960 return -EINVAL;
961 }
962 cur_time = psched_get_time();
963
964 if (tca[TCA_RATE]) {
965 err = gen_replace_estimator(&cl->bstats, NULL,
966 &cl->rate_est,
967 NULL,
968 qdisc_root_sleeping_running(sch),
969 tca[TCA_RATE]);
970 if (err)
971 return err;
972 }
973
974 sch_tree_lock(sch);
975 old_flags = cl->cl_flags;
976
977 if (rsc != NULL)
978 hfsc_change_rsc(cl, rsc, cur_time);
979 if (fsc != NULL)
980 hfsc_change_fsc(cl, fsc);
981 if (usc != NULL)
982 hfsc_change_usc(cl, usc, cur_time);
983
984 if (cl->qdisc->q.qlen != 0) {
985 int len = qdisc_peek_len(cl->qdisc);
986
987 if (cl->cl_flags & HFSC_RSC) {
988 if (old_flags & HFSC_RSC)
989 update_ed(cl, len);
990 else
991 init_ed(cl, len);
992 }
993
994 if (cl->cl_flags & HFSC_FSC) {
995 if (old_flags & HFSC_FSC)
996 update_vf(cl, 0, cur_time);
997 else
998 init_vf(cl, len);
999 }
1000 }
1001 sch_tree_unlock(sch);
1002
1003 return 0;
1004 }
1005
1006 if (parentid == TC_H_ROOT)
1007 return -EEXIST;
1008
1009 parent = &q->root;
1010 if (parentid) {
1011 parent = hfsc_find_class(parentid, sch);
1012 if (parent == NULL)
1013 return -ENOENT;
1014 }
1015
1016 if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1017 return -EINVAL;
1018 if (hfsc_find_class(classid, sch))
1019 return -EEXIST;
1020
1021 if (rsc == NULL && fsc == NULL)
1022 return -EINVAL;
1023
1024 cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1025 if (cl == NULL)
1026 return -ENOBUFS;
1027
1028 err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1029 if (err) {
1030 kfree(cl);
1031 return err;
1032 }
1033
1034 if (tca[TCA_RATE]) {
1035 err = gen_new_estimator(&cl->bstats, NULL, &cl->rate_est,
1036 NULL,
1037 qdisc_root_sleeping_running(sch),
1038 tca[TCA_RATE]);
1039 if (err) {
1040 tcf_block_put(cl->block);
1041 kfree(cl);
1042 return err;
1043 }
1044 }
1045
1046 if (rsc != NULL)
1047 hfsc_change_rsc(cl, rsc, 0);
1048 if (fsc != NULL)
1049 hfsc_change_fsc(cl, fsc);
1050 if (usc != NULL)
1051 hfsc_change_usc(cl, usc, 0);
1052
1053 cl->cl_common.classid = classid;
1054 cl->sched = q;
1055 cl->cl_parent = parent;
1056 cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1057 classid, NULL);
1058 if (cl->qdisc == NULL)
1059 cl->qdisc = &noop_qdisc;
1060 else
1061 qdisc_hash_add(cl->qdisc, true);
1062 INIT_LIST_HEAD(&cl->children);
1063 cl->vt_tree = RB_ROOT;
1064 cl->cf_tree = RB_ROOT;
1065
1066 sch_tree_lock(sch);
1067 qdisc_class_hash_insert(&q->clhash, &cl->cl_common);
1068 list_add_tail(&cl->siblings, &parent->children);
1069 if (parent->level == 0)
1070 qdisc_purge_queue(parent->qdisc);
1071 hfsc_adjust_levels(parent);
1072 sch_tree_unlock(sch);
1073
1074 qdisc_class_hash_grow(sch, &q->clhash);
1075
1076 *arg = (unsigned long)cl;
1077 return 0;
1078}
1079
1080static void
1081hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1082{
1083 struct hfsc_sched *q = qdisc_priv(sch);
1084
1085 tcf_block_put(cl->block);
1086 qdisc_put(cl->qdisc);
1087 gen_kill_estimator(&cl->rate_est);
1088 if (cl != &q->root)
1089 kfree(cl);
1090}
1091
1092static int
1093hfsc_delete_class(struct Qdisc *sch, unsigned long arg,
1094 struct netlink_ext_ack *extack)
1095{
1096 struct hfsc_sched *q = qdisc_priv(sch);
1097 struct hfsc_class *cl = (struct hfsc_class *)arg;
1098
1099 if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1100 return -EBUSY;
1101
1102 sch_tree_lock(sch);
1103
1104 list_del(&cl->siblings);
1105 hfsc_adjust_levels(cl->cl_parent);
1106
1107 qdisc_purge_queue(cl->qdisc);
1108 qdisc_class_hash_remove(&q->clhash, &cl->cl_common);
1109
1110 sch_tree_unlock(sch);
1111
1112 hfsc_destroy_class(sch, cl);
1113 return 0;
1114}
1115
1116static struct hfsc_class *
1117hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1118{
1119 struct hfsc_sched *q = qdisc_priv(sch);
1120 struct hfsc_class *head, *cl;
1121 struct tcf_result res;
1122 struct tcf_proto *tcf;
1123 int result;
1124
1125 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1126 (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1127 if (cl->level == 0)
1128 return cl;
1129
1130 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
1131 head = &q->root;
1132 tcf = rcu_dereference_bh(q->root.filter_list);
1133 while (tcf && (result = tcf_classify(skb, tcf, &res, false)) >= 0) {
1134#ifdef CONFIG_NET_CLS_ACT
1135 switch (result) {
1136 case TC_ACT_QUEUED:
1137 case TC_ACT_STOLEN:
1138 case TC_ACT_TRAP:
1139 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
1140 fallthrough;
1141 case TC_ACT_SHOT:
1142 return NULL;
1143 }
1144#endif
1145 cl = (struct hfsc_class *)res.class;
1146 if (!cl) {
1147 cl = hfsc_find_class(res.classid, sch);
1148 if (!cl)
1149 break; /* filter selected invalid classid */
1150 if (cl->level >= head->level)
1151 break; /* filter may only point downwards */
1152 }
1153
1154 if (cl->level == 0)
1155 return cl; /* hit leaf class */
1156
1157 /* apply inner filter chain */
1158 tcf = rcu_dereference_bh(cl->filter_list);
1159 head = cl;
1160 }
1161
1162 /* classification failed, try default class */
1163 cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1164 if (cl == NULL || cl->level > 0)
1165 return NULL;
1166
1167 return cl;
1168}
1169
1170static int
1171hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1172 struct Qdisc **old, struct netlink_ext_ack *extack)
1173{
1174 struct hfsc_class *cl = (struct hfsc_class *)arg;
1175
1176 if (cl->level > 0)
1177 return -EINVAL;
1178 if (new == NULL) {
1179 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1180 cl->cl_common.classid, NULL);
1181 if (new == NULL)
1182 new = &noop_qdisc;
1183 }
1184
1185 *old = qdisc_replace(sch, new, &cl->qdisc);
1186 return 0;
1187}
1188
1189static struct Qdisc *
1190hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1191{
1192 struct hfsc_class *cl = (struct hfsc_class *)arg;
1193
1194 if (cl->level == 0)
1195 return cl->qdisc;
1196
1197 return NULL;
1198}
1199
1200static void
1201hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1202{
1203 struct hfsc_class *cl = (struct hfsc_class *)arg;
1204
1205 /* vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
1206 * needs to be called explicitly to remove a class from vttree.
1207 */
1208 update_vf(cl, 0, 0);
1209 if (cl->cl_flags & HFSC_RSC)
1210 eltree_remove(cl);
1211}
1212
1213static unsigned long
1214hfsc_search_class(struct Qdisc *sch, u32 classid)
1215{
1216 return (unsigned long)hfsc_find_class(classid, sch);
1217}
1218
1219static unsigned long
1220hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1221{
1222 struct hfsc_class *p = (struct hfsc_class *)parent;
1223 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1224
1225 if (cl != NULL) {
1226 if (p != NULL && p->level <= cl->level)
1227 return 0;
1228 cl->filter_cnt++;
1229 }
1230
1231 return (unsigned long)cl;
1232}
1233
1234static void
1235hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1236{
1237 struct hfsc_class *cl = (struct hfsc_class *)arg;
1238
1239 cl->filter_cnt--;
1240}
1241
1242static struct tcf_block *hfsc_tcf_block(struct Qdisc *sch, unsigned long arg,
1243 struct netlink_ext_ack *extack)
1244{
1245 struct hfsc_sched *q = qdisc_priv(sch);
1246 struct hfsc_class *cl = (struct hfsc_class *)arg;
1247
1248 if (cl == NULL)
1249 cl = &q->root;
1250
1251 return cl->block;
1252}
1253
1254static int
1255hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1256{
1257 struct tc_service_curve tsc;
1258
1259 tsc.m1 = sm2m(sc->sm1);
1260 tsc.d = dx2d(sc->dx);
1261 tsc.m2 = sm2m(sc->sm2);
1262 if (nla_put(skb, attr, sizeof(tsc), &tsc))
1263 goto nla_put_failure;
1264
1265 return skb->len;
1266
1267 nla_put_failure:
1268 return -1;
1269}
1270
1271static int
1272hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1273{
1274 if ((cl->cl_flags & HFSC_RSC) &&
1275 (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1276 goto nla_put_failure;
1277
1278 if ((cl->cl_flags & HFSC_FSC) &&
1279 (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1280 goto nla_put_failure;
1281
1282 if ((cl->cl_flags & HFSC_USC) &&
1283 (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1284 goto nla_put_failure;
1285
1286 return skb->len;
1287
1288 nla_put_failure:
1289 return -1;
1290}
1291
1292static int
1293hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1294 struct tcmsg *tcm)
1295{
1296 struct hfsc_class *cl = (struct hfsc_class *)arg;
1297 struct nlattr *nest;
1298
1299 tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid :
1300 TC_H_ROOT;
1301 tcm->tcm_handle = cl->cl_common.classid;
1302 if (cl->level == 0)
1303 tcm->tcm_info = cl->qdisc->handle;
1304
1305 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1306 if (nest == NULL)
1307 goto nla_put_failure;
1308 if (hfsc_dump_curves(skb, cl) < 0)
1309 goto nla_put_failure;
1310 return nla_nest_end(skb, nest);
1311
1312 nla_put_failure:
1313 nla_nest_cancel(skb, nest);
1314 return -EMSGSIZE;
1315}
1316
1317static int
1318hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1319 struct gnet_dump *d)
1320{
1321 struct hfsc_class *cl = (struct hfsc_class *)arg;
1322 struct tc_hfsc_stats xstats;
1323 __u32 qlen;
1324
1325 qdisc_qstats_qlen_backlog(cl->qdisc, &qlen, &cl->qstats.backlog);
1326 xstats.level = cl->level;
1327 xstats.period = cl->cl_vtperiod;
1328 xstats.work = cl->cl_total;
1329 xstats.rtwork = cl->cl_cumul;
1330
1331 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch), d, NULL, &cl->bstats) < 0 ||
1332 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1333 gnet_stats_copy_queue(d, NULL, &cl->qstats, qlen) < 0)
1334 return -1;
1335
1336 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1337}
1338
1339
1340
1341static void
1342hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1343{
1344 struct hfsc_sched *q = qdisc_priv(sch);
1345 struct hfsc_class *cl;
1346 unsigned int i;
1347
1348 if (arg->stop)
1349 return;
1350
1351 for (i = 0; i < q->clhash.hashsize; i++) {
1352 hlist_for_each_entry(cl, &q->clhash.hash[i],
1353 cl_common.hnode) {
1354 if (arg->count < arg->skip) {
1355 arg->count++;
1356 continue;
1357 }
1358 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1359 arg->stop = 1;
1360 return;
1361 }
1362 arg->count++;
1363 }
1364 }
1365}
1366
1367static void
1368hfsc_schedule_watchdog(struct Qdisc *sch)
1369{
1370 struct hfsc_sched *q = qdisc_priv(sch);
1371 struct hfsc_class *cl;
1372 u64 next_time = 0;
1373
1374 cl = eltree_get_minel(q);
1375 if (cl)
1376 next_time = cl->cl_e;
1377 if (q->root.cl_cfmin != 0) {
1378 if (next_time == 0 || next_time > q->root.cl_cfmin)
1379 next_time = q->root.cl_cfmin;
1380 }
1381 if (next_time)
1382 qdisc_watchdog_schedule(&q->watchdog, next_time);
1383}
1384
1385static int
1386hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1387 struct netlink_ext_ack *extack)
1388{
1389 struct hfsc_sched *q = qdisc_priv(sch);
1390 struct tc_hfsc_qopt *qopt;
1391 int err;
1392
1393 qdisc_watchdog_init(&q->watchdog, sch);
1394
1395 if (!opt || nla_len(opt) < sizeof(*qopt))
1396 return -EINVAL;
1397 qopt = nla_data(opt);
1398
1399 q->defcls = qopt->defcls;
1400 err = qdisc_class_hash_init(&q->clhash);
1401 if (err < 0)
1402 return err;
1403 q->eligible = RB_ROOT;
1404
1405 err = tcf_block_get(&q->root.block, &q->root.filter_list, sch, extack);
1406 if (err)
1407 return err;
1408
1409 q->root.cl_common.classid = sch->handle;
1410 q->root.sched = q;
1411 q->root.qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1412 sch->handle, NULL);
1413 if (q->root.qdisc == NULL)
1414 q->root.qdisc = &noop_qdisc;
1415 else
1416 qdisc_hash_add(q->root.qdisc, true);
1417 INIT_LIST_HEAD(&q->root.children);
1418 q->root.vt_tree = RB_ROOT;
1419 q->root.cf_tree = RB_ROOT;
1420
1421 qdisc_class_hash_insert(&q->clhash, &q->root.cl_common);
1422 qdisc_class_hash_grow(sch, &q->clhash);
1423
1424 return 0;
1425}
1426
1427static int
1428hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt,
1429 struct netlink_ext_ack *extack)
1430{
1431 struct hfsc_sched *q = qdisc_priv(sch);
1432 struct tc_hfsc_qopt *qopt;
1433
1434 if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1435 return -EINVAL;
1436 qopt = nla_data(opt);
1437
1438 sch_tree_lock(sch);
1439 q->defcls = qopt->defcls;
1440 sch_tree_unlock(sch);
1441
1442 return 0;
1443}
1444
1445static void
1446hfsc_reset_class(struct hfsc_class *cl)
1447{
1448 cl->cl_total = 0;
1449 cl->cl_cumul = 0;
1450 cl->cl_d = 0;
1451 cl->cl_e = 0;
1452 cl->cl_vt = 0;
1453 cl->cl_vtadj = 0;
1454 cl->cl_cvtmin = 0;
1455 cl->cl_cvtoff = 0;
1456 cl->cl_vtperiod = 0;
1457 cl->cl_parentperiod = 0;
1458 cl->cl_f = 0;
1459 cl->cl_myf = 0;
1460 cl->cl_cfmin = 0;
1461 cl->cl_nactive = 0;
1462
1463 cl->vt_tree = RB_ROOT;
1464 cl->cf_tree = RB_ROOT;
1465 qdisc_reset(cl->qdisc);
1466
1467 if (cl->cl_flags & HFSC_RSC)
1468 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1469 if (cl->cl_flags & HFSC_FSC)
1470 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1471 if (cl->cl_flags & HFSC_USC)
1472 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1473}
1474
1475static void
1476hfsc_reset_qdisc(struct Qdisc *sch)
1477{
1478 struct hfsc_sched *q = qdisc_priv(sch);
1479 struct hfsc_class *cl;
1480 unsigned int i;
1481
1482 for (i = 0; i < q->clhash.hashsize; i++) {
1483 hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1484 hfsc_reset_class(cl);
1485 }
1486 q->eligible = RB_ROOT;
1487 qdisc_watchdog_cancel(&q->watchdog);
1488 sch->qstats.backlog = 0;
1489 sch->q.qlen = 0;
1490}
1491
1492static void
1493hfsc_destroy_qdisc(struct Qdisc *sch)
1494{
1495 struct hfsc_sched *q = qdisc_priv(sch);
1496 struct hlist_node *next;
1497 struct hfsc_class *cl;
1498 unsigned int i;
1499
1500 for (i = 0; i < q->clhash.hashsize; i++) {
1501 hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode) {
1502 tcf_block_put(cl->block);
1503 cl->block = NULL;
1504 }
1505 }
1506 for (i = 0; i < q->clhash.hashsize; i++) {
1507 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1508 cl_common.hnode)
1509 hfsc_destroy_class(sch, cl);
1510 }
1511 qdisc_class_hash_destroy(&q->clhash);
1512 qdisc_watchdog_cancel(&q->watchdog);
1513}
1514
1515static int
1516hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1517{
1518 struct hfsc_sched *q = qdisc_priv(sch);
1519 unsigned char *b = skb_tail_pointer(skb);
1520 struct tc_hfsc_qopt qopt;
1521
1522 qopt.defcls = q->defcls;
1523 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1524 goto nla_put_failure;
1525 return skb->len;
1526
1527 nla_put_failure:
1528 nlmsg_trim(skb, b);
1529 return -1;
1530}
1531
1532static int
1533hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free)
1534{
1535 unsigned int len = qdisc_pkt_len(skb);
1536 struct hfsc_class *cl;
1537 int err;
1538 bool first;
1539
1540 cl = hfsc_classify(skb, sch, &err);
1541 if (cl == NULL) {
1542 if (err & __NET_XMIT_BYPASS)
1543 qdisc_qstats_drop(sch);
1544 __qdisc_drop(skb, to_free);
1545 return err;
1546 }
1547
1548 first = !cl->qdisc->q.qlen;
1549 err = qdisc_enqueue(skb, cl->qdisc, to_free);
1550 if (unlikely(err != NET_XMIT_SUCCESS)) {
1551 if (net_xmit_drop_count(err)) {
1552 cl->qstats.drops++;
1553 qdisc_qstats_drop(sch);
1554 }
1555 return err;
1556 }
1557
1558 if (first) {
1559 if (cl->cl_flags & HFSC_RSC)
1560 init_ed(cl, len);
1561 if (cl->cl_flags & HFSC_FSC)
1562 init_vf(cl, len);
1563 /*
1564 * If this is the first packet, isolate the head so an eventual
1565 * head drop before the first dequeue operation has no chance
1566 * to invalidate the deadline.
1567 */
1568 if (cl->cl_flags & HFSC_RSC)
1569 cl->qdisc->ops->peek(cl->qdisc);
1570
1571 }
1572
1573 sch->qstats.backlog += len;
1574 sch->q.qlen++;
1575
1576 return NET_XMIT_SUCCESS;
1577}
1578
1579static struct sk_buff *
1580hfsc_dequeue(struct Qdisc *sch)
1581{
1582 struct hfsc_sched *q = qdisc_priv(sch);
1583 struct hfsc_class *cl;
1584 struct sk_buff *skb;
1585 u64 cur_time;
1586 unsigned int next_len;
1587 int realtime = 0;
1588
1589 if (sch->q.qlen == 0)
1590 return NULL;
1591
1592 cur_time = psched_get_time();
1593
1594 /*
1595 * if there are eligible classes, use real-time criteria.
1596 * find the class with the minimum deadline among
1597 * the eligible classes.
1598 */
1599 cl = eltree_get_mindl(q, cur_time);
1600 if (cl) {
1601 realtime = 1;
1602 } else {
1603 /*
1604 * use link-sharing criteria
1605 * get the class with the minimum vt in the hierarchy
1606 */
1607 cl = vttree_get_minvt(&q->root, cur_time);
1608 if (cl == NULL) {
1609 qdisc_qstats_overlimit(sch);
1610 hfsc_schedule_watchdog(sch);
1611 return NULL;
1612 }
1613 }
1614
1615 skb = qdisc_dequeue_peeked(cl->qdisc);
1616 if (skb == NULL) {
1617 qdisc_warn_nonwc("HFSC", cl->qdisc);
1618 return NULL;
1619 }
1620
1621 bstats_update(&cl->bstats, skb);
1622 update_vf(cl, qdisc_pkt_len(skb), cur_time);
1623 if (realtime)
1624 cl->cl_cumul += qdisc_pkt_len(skb);
1625
1626 if (cl->cl_flags & HFSC_RSC) {
1627 if (cl->qdisc->q.qlen != 0) {
1628 /* update ed */
1629 next_len = qdisc_peek_len(cl->qdisc);
1630 if (realtime)
1631 update_ed(cl, next_len);
1632 else
1633 update_d(cl, next_len);
1634 } else {
1635 /* the class becomes passive */
1636 eltree_remove(cl);
1637 }
1638 }
1639
1640 qdisc_bstats_update(sch, skb);
1641 qdisc_qstats_backlog_dec(sch, skb);
1642 sch->q.qlen--;
1643
1644 return skb;
1645}
1646
1647static const struct Qdisc_class_ops hfsc_class_ops = {
1648 .change = hfsc_change_class,
1649 .delete = hfsc_delete_class,
1650 .graft = hfsc_graft_class,
1651 .leaf = hfsc_class_leaf,
1652 .qlen_notify = hfsc_qlen_notify,
1653 .find = hfsc_search_class,
1654 .bind_tcf = hfsc_bind_tcf,
1655 .unbind_tcf = hfsc_unbind_tcf,
1656 .tcf_block = hfsc_tcf_block,
1657 .dump = hfsc_dump_class,
1658 .dump_stats = hfsc_dump_class_stats,
1659 .walk = hfsc_walk
1660};
1661
1662static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1663 .id = "hfsc",
1664 .init = hfsc_init_qdisc,
1665 .change = hfsc_change_qdisc,
1666 .reset = hfsc_reset_qdisc,
1667 .destroy = hfsc_destroy_qdisc,
1668 .dump = hfsc_dump_qdisc,
1669 .enqueue = hfsc_enqueue,
1670 .dequeue = hfsc_dequeue,
1671 .peek = qdisc_peek_dequeued,
1672 .cl_ops = &hfsc_class_ops,
1673 .priv_size = sizeof(struct hfsc_sched),
1674 .owner = THIS_MODULE
1675};
1676
1677static int __init
1678hfsc_init(void)
1679{
1680 return register_qdisc(&hfsc_qdisc_ops);
1681}
1682
1683static void __exit
1684hfsc_cleanup(void)
1685{
1686 unregister_qdisc(&hfsc_qdisc_ops);
1687}
1688
1689MODULE_LICENSE("GPL");
1690module_init(hfsc_init);
1691module_exit(hfsc_cleanup);