Merge tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost
[cris-mirror.git] / net / sched / sch_hfsc.c
blob3ae9877ea2057d0ba517c84d38f6ba6a79ff6ef8
1 /*
2 * Copyright (c) 2003 Patrick McHardy, <kaber@trash.net>
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.
9 * 2003-10-17 - Ported from altq
12 * Copyright (c) 1997-1999 Carnegie Mellon University. All Rights Reserved.
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.
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.
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.
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.
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.
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>
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.
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.
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 */
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 */
105 enum hfsc_class_flags {
106 HFSC_RSC = 0x1,
107 HFSC_FSC = 0x2,
108 HFSC_USC = 0x4
111 struct hfsc_class {
112 struct Qdisc_class_common cl_common;
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 */
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 */
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 */
134 u64 cl_total; /* total work in bytes */
135 u64 cl_cumul; /* cumulative work in bytes done by
136 real-time criteria */
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 */
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 */
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 */
169 struct 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 */
177 #define HT_INFINITY 0xffffffffffffffffULL /* infinite time value */
181 * eligible tree holds backlogged classes being sorted by their eligible times.
182 * there is one eligible tree per hfsc instance.
185 static void
186 eltree_insert(struct hfsc_class *cl)
188 struct rb_node **p = &cl->sched->eligible.rb_node;
189 struct rb_node *parent = NULL;
190 struct hfsc_class *cl1;
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;
200 rb_link_node(&cl->el_node, parent, p);
201 rb_insert_color(&cl->el_node, &cl->sched->eligible);
204 static inline void
205 eltree_remove(struct hfsc_class *cl)
207 rb_erase(&cl->el_node, &cl->sched->eligible);
210 static inline void
211 eltree_update(struct hfsc_class *cl)
213 eltree_remove(cl);
214 eltree_insert(cl);
217 /* find the class with the minimum deadline among the eligible classes */
218 static inline struct hfsc_class *
219 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
221 struct hfsc_class *p, *cl = NULL;
222 struct rb_node *n;
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;
231 return cl;
234 /* find the class with minimum eligible time among the eligible classes */
235 static inline struct hfsc_class *
236 eltree_get_minel(struct hfsc_sched *q)
238 struct rb_node *n;
240 n = rb_first(&q->eligible);
241 if (n == NULL)
242 return NULL;
243 return rb_entry(n, struct hfsc_class, el_node);
247 * vttree holds holds backlogged child classes being sorted by their virtual
248 * time. each intermediate class has one vttree.
250 static void
251 vttree_insert(struct hfsc_class *cl)
253 struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
254 struct rb_node *parent = NULL;
255 struct hfsc_class *cl1;
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;
265 rb_link_node(&cl->vt_node, parent, p);
266 rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
269 static inline void
270 vttree_remove(struct hfsc_class *cl)
272 rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
275 static inline void
276 vttree_update(struct hfsc_class *cl)
278 vttree_remove(cl);
279 vttree_insert(cl);
282 static inline struct hfsc_class *
283 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
285 struct hfsc_class *p;
286 struct rb_node *n;
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;
293 return NULL;
297 * get the leaf class with the minimum vt in the hierarchy
299 static struct hfsc_class *
300 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
302 /* if root-class's cfmin is bigger than cur_time nothing to do */
303 if (cl->cl_cfmin > cur_time)
304 return NULL;
306 while (cl->level > 0) {
307 cl = vttree_firstfit(cl, cur_time);
308 if (cl == NULL)
309 return NULL;
311 * update parent's cl_cvtmin.
313 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
314 cl->cl_parent->cl_cvtmin = cl->cl_vt;
316 return cl;
319 static void
320 cftree_insert(struct hfsc_class *cl)
322 struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
323 struct rb_node *parent = NULL;
324 struct hfsc_class *cl1;
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;
334 rb_link_node(&cl->cf_node, parent, p);
335 rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
338 static inline void
339 cftree_remove(struct hfsc_class *cl)
341 rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
344 static inline void
345 cftree_update(struct hfsc_class *cl)
347 cftree_remove(cl);
348 cftree_insert(cl);
352 * service curve support functions
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
362 * The clock source resolution with ktime and PSCHED_SHIFT 10 is 1.024us.
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.
368 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
369 * ------------+-------------------------------------------------------
370 * bytes/1.024us 12.8e-3 128e-3 1280e-3 12800e-3 128000e-3
372 * 1.024us/byte 78.125 7.8125 0.78125 0.078125 0.0078125
374 * So, for PSCHED_SHIFT 10 we need: SM_SHIFT 20, ISM_SHIFT 18.
376 #define SM_SHIFT (30 - PSCHED_SHIFT)
377 #define ISM_SHIFT (8 + PSCHED_SHIFT)
379 #define SM_MASK ((1ULL << SM_SHIFT) - 1)
380 #define ISM_MASK ((1ULL << ISM_SHIFT) - 1)
382 static inline u64
383 seg_x2y(u64 x, u64 sm)
385 u64 y;
388 * compute
389 * y = x * sm >> SM_SHIFT
390 * but divide it for the upper and lower bits to avoid overflow
392 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
393 return y;
396 static inline u64
397 seg_y2x(u64 y, u64 ism)
399 u64 x;
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);
409 return x;
412 /* Convert m (bps) into sm (bytes/psched us) */
413 static u64
414 m2sm(u32 m)
416 u64 sm;
418 sm = ((u64)m << SM_SHIFT);
419 sm += PSCHED_TICKS_PER_SEC - 1;
420 do_div(sm, PSCHED_TICKS_PER_SEC);
421 return sm;
424 /* convert m (bps) into ism (psched us/byte) */
425 static u64
426 m2ism(u32 m)
428 u64 ism;
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);
437 return ism;
440 /* convert d (us) into dx (psched us) */
441 static u64
442 d2dx(u32 d)
444 u64 dx;
446 dx = ((u64)d * PSCHED_TICKS_PER_SEC);
447 dx += USEC_PER_SEC - 1;
448 do_div(dx, USEC_PER_SEC);
449 return dx;
452 /* convert sm (bytes/psched us) into m (bps) */
453 static u32
454 sm2m(u64 sm)
456 u64 m;
458 m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
459 return (u32)m;
462 /* convert dx (psched us) into d (us) */
463 static u32
464 dx2d(u64 dx)
466 u64 d;
468 d = dx * USEC_PER_SEC;
469 do_div(d, PSCHED_TICKS_PER_SEC);
470 return (u32)d;
473 static void
474 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
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);
485 * initialize the runtime service curve with the given internal
486 * service curve starting at (x, y).
488 static void
489 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
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;
502 * calculate the y-projection of the runtime service curve by the
503 * given x-projection value
505 static u64
506 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
508 u64 x;
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);
523 return x;
526 static u64
527 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
529 u64 y;
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;
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).
547 static void
548 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
550 u64 y1, y2, dx, dy;
551 u32 dsm;
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;
565 * service curve is concave
566 * compute the two y values of the current rtsc
567 * y1: at x
568 * y2: at (x + dx)
570 y1 = rtsc_x2y(rtsc, x);
571 if (y1 <= y) {
572 /* rtsc is below isc, no change to rtsc */
573 return;
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;
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)
592 dx = (y1 - y) << SM_SHIFT;
593 dsm = isc->sm1 - isc->sm2;
594 do_div(dx, dsm);
596 * check if (x, y1) belongs to the 1st segment of rtsc.
597 * if so, add the offset.
599 if (rtsc->x + rtsc->dx > x)
600 dx += rtsc->x + rtsc->dx - x;
601 dy = seg_x2y(dx, isc->sm1);
603 rtsc->x = x;
604 rtsc->y = y;
605 rtsc->dx = dx;
606 rtsc->dy = dy;
609 static void
610 init_ed(struct hfsc_class *cl, unsigned int next_len)
612 u64 cur_time = psched_get_time();
614 /* update the deadline curve */
615 rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
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.
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;
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);
632 eltree_insert(cl);
635 static void
636 update_ed(struct hfsc_class *cl, unsigned int next_len)
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);
641 eltree_update(cl);
644 static inline void
645 update_d(struct hfsc_class *cl, unsigned int next_len)
647 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
650 static inline void
651 update_cfmin(struct hfsc_class *cl)
653 struct rb_node *n = rb_first(&cl->cf_tree);
654 struct hfsc_class *p;
656 if (n == NULL) {
657 cl->cl_cfmin = 0;
658 return;
660 p = rb_entry(n, struct hfsc_class, cf_node);
661 cl->cl_cfmin = p->cl_f;
664 static void
665 init_vf(struct hfsc_class *cl, unsigned int len)
667 struct hfsc_class *max_cl;
668 struct rb_node *n;
669 u64 vt, f, cur_time;
670 int go_active;
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;
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);
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.
689 vt = max_cl->cl_vt;
690 if (cl->cl_parent->cl_cvtmin != 0)
691 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
693 if (cl->cl_parent->cl_vtperiod !=
694 cl->cl_parentperiod || vt > cl->cl_vt)
695 cl->cl_vt = vt;
696 } else {
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.
703 cl->cl_vt = cl->cl_parent->cl_cvtoff;
704 cl->cl_parent->cl_cvtmin = 0;
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;
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;
717 vttree_insert(cl);
718 cftree_insert(cl);
720 if (cl->cl_flags & HFSC_USC) {
721 /* class has upper limit curve */
722 if (cur_time == 0)
723 cur_time = psched_get_time();
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);
734 f = max(cl->cl_myf, cl->cl_cfmin);
735 if (f != cl->cl_f) {
736 cl->cl_f = f;
737 cftree_update(cl);
739 update_cfmin(cl->cl_parent);
743 static void
744 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
746 u64 f; /* , myf_bound, delta; */
747 int go_passive = 0;
749 if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
750 go_passive = 1;
752 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
753 cl->cl_total += len;
755 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
756 continue;
758 if (go_passive && --cl->cl_nactive == 0)
759 go_passive = 1;
760 else
761 go_passive = 0;
763 /* update vt */
764 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total) + cl->cl_vtadj;
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.
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;
776 if (go_passive) {
777 /* no more active child, going passive */
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;
783 /* remove this class from the vt tree */
784 vttree_remove(cl);
786 cftree_remove(cl);
787 update_cfmin(cl->cl_parent);
789 continue;
792 /* update the vt tree */
793 vttree_update(cl);
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);
802 * This code causes classes to stay way under their
803 * limit when multiple classes are used at gigabit
804 * speed. needs investigation. -kaber
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.
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;
819 #endif
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);
831 static unsigned int
832 qdisc_peek_len(struct Qdisc *sch)
834 struct sk_buff *skb;
835 unsigned int len;
837 skb = sch->ops->peek(sch);
838 if (unlikely(skb == NULL)) {
839 qdisc_warn_nonwc("qdisc_peek_len", sch);
840 return 0;
842 len = qdisc_pkt_len(skb);
844 return len;
847 static void
848 hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl)
850 unsigned int len = cl->qdisc->q.qlen;
851 unsigned int backlog = cl->qdisc->qstats.backlog;
853 qdisc_reset(cl->qdisc);
854 qdisc_tree_reduce_backlog(cl->qdisc, len, backlog);
857 static void
858 hfsc_adjust_levels(struct hfsc_class *cl)
860 struct hfsc_class *p;
861 unsigned int level;
863 do {
864 level = 0;
865 list_for_each_entry(p, &cl->children, siblings) {
866 if (p->level >= level)
867 level = p->level + 1;
869 cl->level = level;
870 } while ((cl = cl->cl_parent) != NULL);
873 static inline struct hfsc_class *
874 hfsc_find_class(u32 classid, struct Qdisc *sch)
876 struct hfsc_sched *q = qdisc_priv(sch);
877 struct Qdisc_class_common *clc;
879 clc = qdisc_class_find(&q->clhash, classid);
880 if (clc == NULL)
881 return NULL;
882 return container_of(clc, struct hfsc_class, cl_common);
885 static void
886 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
887 u64 cur_time)
889 sc2isc(rsc, &cl->cl_rsc);
890 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
891 cl->cl_eligible = cl->cl_deadline;
892 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
893 cl->cl_eligible.dx = 0;
894 cl->cl_eligible.dy = 0;
896 cl->cl_flags |= HFSC_RSC;
899 static void
900 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
902 sc2isc(fsc, &cl->cl_fsc);
903 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
904 cl->cl_flags |= HFSC_FSC;
907 static void
908 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
909 u64 cur_time)
911 sc2isc(usc, &cl->cl_usc);
912 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
913 cl->cl_flags |= HFSC_USC;
916 static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
917 [TCA_HFSC_RSC] = { .len = sizeof(struct tc_service_curve) },
918 [TCA_HFSC_FSC] = { .len = sizeof(struct tc_service_curve) },
919 [TCA_HFSC_USC] = { .len = sizeof(struct tc_service_curve) },
922 static int
923 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
924 struct nlattr **tca, unsigned long *arg,
925 struct netlink_ext_ack *extack)
927 struct hfsc_sched *q = qdisc_priv(sch);
928 struct hfsc_class *cl = (struct hfsc_class *)*arg;
929 struct hfsc_class *parent = NULL;
930 struct nlattr *opt = tca[TCA_OPTIONS];
931 struct nlattr *tb[TCA_HFSC_MAX + 1];
932 struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
933 u64 cur_time;
934 int err;
936 if (opt == NULL)
937 return -EINVAL;
939 err = nla_parse_nested(tb, TCA_HFSC_MAX, opt, hfsc_policy, NULL);
940 if (err < 0)
941 return err;
943 if (tb[TCA_HFSC_RSC]) {
944 rsc = nla_data(tb[TCA_HFSC_RSC]);
945 if (rsc->m1 == 0 && rsc->m2 == 0)
946 rsc = NULL;
949 if (tb[TCA_HFSC_FSC]) {
950 fsc = nla_data(tb[TCA_HFSC_FSC]);
951 if (fsc->m1 == 0 && fsc->m2 == 0)
952 fsc = NULL;
955 if (tb[TCA_HFSC_USC]) {
956 usc = nla_data(tb[TCA_HFSC_USC]);
957 if (usc->m1 == 0 && usc->m2 == 0)
958 usc = NULL;
961 if (cl != NULL) {
962 int old_flags;
964 if (parentid) {
965 if (cl->cl_parent &&
966 cl->cl_parent->cl_common.classid != parentid)
967 return -EINVAL;
968 if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
969 return -EINVAL;
971 cur_time = psched_get_time();
973 if (tca[TCA_RATE]) {
974 err = gen_replace_estimator(&cl->bstats, NULL,
975 &cl->rate_est,
976 NULL,
977 qdisc_root_sleeping_running(sch),
978 tca[TCA_RATE]);
979 if (err)
980 return err;
983 sch_tree_lock(sch);
984 old_flags = cl->cl_flags;
986 if (rsc != NULL)
987 hfsc_change_rsc(cl, rsc, cur_time);
988 if (fsc != NULL)
989 hfsc_change_fsc(cl, fsc);
990 if (usc != NULL)
991 hfsc_change_usc(cl, usc, cur_time);
993 if (cl->qdisc->q.qlen != 0) {
994 int len = qdisc_peek_len(cl->qdisc);
996 if (cl->cl_flags & HFSC_RSC) {
997 if (old_flags & HFSC_RSC)
998 update_ed(cl, len);
999 else
1000 init_ed(cl, len);
1003 if (cl->cl_flags & HFSC_FSC) {
1004 if (old_flags & HFSC_FSC)
1005 update_vf(cl, 0, cur_time);
1006 else
1007 init_vf(cl, len);
1010 sch_tree_unlock(sch);
1012 return 0;
1015 if (parentid == TC_H_ROOT)
1016 return -EEXIST;
1018 parent = &q->root;
1019 if (parentid) {
1020 parent = hfsc_find_class(parentid, sch);
1021 if (parent == NULL)
1022 return -ENOENT;
1025 if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1026 return -EINVAL;
1027 if (hfsc_find_class(classid, sch))
1028 return -EEXIST;
1030 if (rsc == NULL && fsc == NULL)
1031 return -EINVAL;
1033 cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1034 if (cl == NULL)
1035 return -ENOBUFS;
1037 err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1038 if (err) {
1039 kfree(cl);
1040 return err;
1043 if (tca[TCA_RATE]) {
1044 err = gen_new_estimator(&cl->bstats, NULL, &cl->rate_est,
1045 NULL,
1046 qdisc_root_sleeping_running(sch),
1047 tca[TCA_RATE]);
1048 if (err) {
1049 tcf_block_put(cl->block);
1050 kfree(cl);
1051 return err;
1055 if (rsc != NULL)
1056 hfsc_change_rsc(cl, rsc, 0);
1057 if (fsc != NULL)
1058 hfsc_change_fsc(cl, fsc);
1059 if (usc != NULL)
1060 hfsc_change_usc(cl, usc, 0);
1062 cl->cl_common.classid = classid;
1063 cl->sched = q;
1064 cl->cl_parent = parent;
1065 cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1066 classid, NULL);
1067 if (cl->qdisc == NULL)
1068 cl->qdisc = &noop_qdisc;
1069 else
1070 qdisc_hash_add(cl->qdisc, true);
1071 INIT_LIST_HEAD(&cl->children);
1072 cl->vt_tree = RB_ROOT;
1073 cl->cf_tree = RB_ROOT;
1075 sch_tree_lock(sch);
1076 qdisc_class_hash_insert(&q->clhash, &cl->cl_common);
1077 list_add_tail(&cl->siblings, &parent->children);
1078 if (parent->level == 0)
1079 hfsc_purge_queue(sch, parent);
1080 hfsc_adjust_levels(parent);
1081 sch_tree_unlock(sch);
1083 qdisc_class_hash_grow(sch, &q->clhash);
1085 *arg = (unsigned long)cl;
1086 return 0;
1089 static void
1090 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1092 struct hfsc_sched *q = qdisc_priv(sch);
1094 tcf_block_put(cl->block);
1095 qdisc_destroy(cl->qdisc);
1096 gen_kill_estimator(&cl->rate_est);
1097 if (cl != &q->root)
1098 kfree(cl);
1101 static int
1102 hfsc_delete_class(struct Qdisc *sch, unsigned long arg)
1104 struct hfsc_sched *q = qdisc_priv(sch);
1105 struct hfsc_class *cl = (struct hfsc_class *)arg;
1107 if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1108 return -EBUSY;
1110 sch_tree_lock(sch);
1112 list_del(&cl->siblings);
1113 hfsc_adjust_levels(cl->cl_parent);
1115 hfsc_purge_queue(sch, cl);
1116 qdisc_class_hash_remove(&q->clhash, &cl->cl_common);
1118 sch_tree_unlock(sch);
1120 hfsc_destroy_class(sch, cl);
1121 return 0;
1124 static struct hfsc_class *
1125 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1127 struct hfsc_sched *q = qdisc_priv(sch);
1128 struct hfsc_class *head, *cl;
1129 struct tcf_result res;
1130 struct tcf_proto *tcf;
1131 int result;
1133 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1134 (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1135 if (cl->level == 0)
1136 return cl;
1138 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
1139 head = &q->root;
1140 tcf = rcu_dereference_bh(q->root.filter_list);
1141 while (tcf && (result = tcf_classify(skb, tcf, &res, false)) >= 0) {
1142 #ifdef CONFIG_NET_CLS_ACT
1143 switch (result) {
1144 case TC_ACT_QUEUED:
1145 case TC_ACT_STOLEN:
1146 case TC_ACT_TRAP:
1147 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
1148 /* fall through */
1149 case TC_ACT_SHOT:
1150 return NULL;
1152 #endif
1153 cl = (struct hfsc_class *)res.class;
1154 if (!cl) {
1155 cl = hfsc_find_class(res.classid, sch);
1156 if (!cl)
1157 break; /* filter selected invalid classid */
1158 if (cl->level >= head->level)
1159 break; /* filter may only point downwards */
1162 if (cl->level == 0)
1163 return cl; /* hit leaf class */
1165 /* apply inner filter chain */
1166 tcf = rcu_dereference_bh(cl->filter_list);
1167 head = cl;
1170 /* classification failed, try default class */
1171 cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1172 if (cl == NULL || cl->level > 0)
1173 return NULL;
1175 return cl;
1178 static int
1179 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1180 struct Qdisc **old, struct netlink_ext_ack *extack)
1182 struct hfsc_class *cl = (struct hfsc_class *)arg;
1184 if (cl->level > 0)
1185 return -EINVAL;
1186 if (new == NULL) {
1187 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1188 cl->cl_common.classid, NULL);
1189 if (new == NULL)
1190 new = &noop_qdisc;
1193 *old = qdisc_replace(sch, new, &cl->qdisc);
1194 return 0;
1197 static struct Qdisc *
1198 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1200 struct hfsc_class *cl = (struct hfsc_class *)arg;
1202 if (cl->level == 0)
1203 return cl->qdisc;
1205 return NULL;
1208 static void
1209 hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1211 struct hfsc_class *cl = (struct hfsc_class *)arg;
1213 /* vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
1214 * needs to be called explicitly to remove a class from vttree.
1216 update_vf(cl, 0, 0);
1217 if (cl->cl_flags & HFSC_RSC)
1218 eltree_remove(cl);
1221 static unsigned long
1222 hfsc_search_class(struct Qdisc *sch, u32 classid)
1224 return (unsigned long)hfsc_find_class(classid, sch);
1227 static unsigned long
1228 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1230 struct hfsc_class *p = (struct hfsc_class *)parent;
1231 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1233 if (cl != NULL) {
1234 if (p != NULL && p->level <= cl->level)
1235 return 0;
1236 cl->filter_cnt++;
1239 return (unsigned long)cl;
1242 static void
1243 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1245 struct hfsc_class *cl = (struct hfsc_class *)arg;
1247 cl->filter_cnt--;
1250 static struct tcf_block *hfsc_tcf_block(struct Qdisc *sch, unsigned long arg,
1251 struct netlink_ext_ack *extack)
1253 struct hfsc_sched *q = qdisc_priv(sch);
1254 struct hfsc_class *cl = (struct hfsc_class *)arg;
1256 if (cl == NULL)
1257 cl = &q->root;
1259 return cl->block;
1262 static int
1263 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1265 struct tc_service_curve tsc;
1267 tsc.m1 = sm2m(sc->sm1);
1268 tsc.d = dx2d(sc->dx);
1269 tsc.m2 = sm2m(sc->sm2);
1270 if (nla_put(skb, attr, sizeof(tsc), &tsc))
1271 goto nla_put_failure;
1273 return skb->len;
1275 nla_put_failure:
1276 return -1;
1279 static int
1280 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1282 if ((cl->cl_flags & HFSC_RSC) &&
1283 (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1284 goto nla_put_failure;
1286 if ((cl->cl_flags & HFSC_FSC) &&
1287 (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1288 goto nla_put_failure;
1290 if ((cl->cl_flags & HFSC_USC) &&
1291 (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1292 goto nla_put_failure;
1294 return skb->len;
1296 nla_put_failure:
1297 return -1;
1300 static int
1301 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1302 struct tcmsg *tcm)
1304 struct hfsc_class *cl = (struct hfsc_class *)arg;
1305 struct nlattr *nest;
1307 tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid :
1308 TC_H_ROOT;
1309 tcm->tcm_handle = cl->cl_common.classid;
1310 if (cl->level == 0)
1311 tcm->tcm_info = cl->qdisc->handle;
1313 nest = nla_nest_start(skb, TCA_OPTIONS);
1314 if (nest == NULL)
1315 goto nla_put_failure;
1316 if (hfsc_dump_curves(skb, cl) < 0)
1317 goto nla_put_failure;
1318 return nla_nest_end(skb, nest);
1320 nla_put_failure:
1321 nla_nest_cancel(skb, nest);
1322 return -EMSGSIZE;
1325 static int
1326 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1327 struct gnet_dump *d)
1329 struct hfsc_class *cl = (struct hfsc_class *)arg;
1330 struct tc_hfsc_stats xstats;
1332 cl->qstats.backlog = cl->qdisc->qstats.backlog;
1333 xstats.level = cl->level;
1334 xstats.period = cl->cl_vtperiod;
1335 xstats.work = cl->cl_total;
1336 xstats.rtwork = cl->cl_cumul;
1338 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch), d, NULL, &cl->bstats) < 0 ||
1339 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1340 gnet_stats_copy_queue(d, NULL, &cl->qstats, cl->qdisc->q.qlen) < 0)
1341 return -1;
1343 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1348 static void
1349 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1351 struct hfsc_sched *q = qdisc_priv(sch);
1352 struct hfsc_class *cl;
1353 unsigned int i;
1355 if (arg->stop)
1356 return;
1358 for (i = 0; i < q->clhash.hashsize; i++) {
1359 hlist_for_each_entry(cl, &q->clhash.hash[i],
1360 cl_common.hnode) {
1361 if (arg->count < arg->skip) {
1362 arg->count++;
1363 continue;
1365 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1366 arg->stop = 1;
1367 return;
1369 arg->count++;
1374 static void
1375 hfsc_schedule_watchdog(struct Qdisc *sch)
1377 struct hfsc_sched *q = qdisc_priv(sch);
1378 struct hfsc_class *cl;
1379 u64 next_time = 0;
1381 cl = eltree_get_minel(q);
1382 if (cl)
1383 next_time = cl->cl_e;
1384 if (q->root.cl_cfmin != 0) {
1385 if (next_time == 0 || next_time > q->root.cl_cfmin)
1386 next_time = q->root.cl_cfmin;
1388 WARN_ON(next_time == 0);
1389 qdisc_watchdog_schedule(&q->watchdog, next_time);
1392 static int
1393 hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1394 struct netlink_ext_ack *extack)
1396 struct hfsc_sched *q = qdisc_priv(sch);
1397 struct tc_hfsc_qopt *qopt;
1398 int err;
1400 qdisc_watchdog_init(&q->watchdog, sch);
1402 if (!opt || nla_len(opt) < sizeof(*qopt))
1403 return -EINVAL;
1404 qopt = nla_data(opt);
1406 q->defcls = qopt->defcls;
1407 err = qdisc_class_hash_init(&q->clhash);
1408 if (err < 0)
1409 return err;
1410 q->eligible = RB_ROOT;
1412 err = tcf_block_get(&q->root.block, &q->root.filter_list, sch, extack);
1413 if (err)
1414 return err;
1416 q->root.cl_common.classid = sch->handle;
1417 q->root.sched = q;
1418 q->root.qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1419 sch->handle, NULL);
1420 if (q->root.qdisc == NULL)
1421 q->root.qdisc = &noop_qdisc;
1422 else
1423 qdisc_hash_add(q->root.qdisc, true);
1424 INIT_LIST_HEAD(&q->root.children);
1425 q->root.vt_tree = RB_ROOT;
1426 q->root.cf_tree = RB_ROOT;
1428 qdisc_class_hash_insert(&q->clhash, &q->root.cl_common);
1429 qdisc_class_hash_grow(sch, &q->clhash);
1431 return 0;
1434 static int
1435 hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt,
1436 struct netlink_ext_ack *extack)
1438 struct hfsc_sched *q = qdisc_priv(sch);
1439 struct tc_hfsc_qopt *qopt;
1441 if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1442 return -EINVAL;
1443 qopt = nla_data(opt);
1445 sch_tree_lock(sch);
1446 q->defcls = qopt->defcls;
1447 sch_tree_unlock(sch);
1449 return 0;
1452 static void
1453 hfsc_reset_class(struct hfsc_class *cl)
1455 cl->cl_total = 0;
1456 cl->cl_cumul = 0;
1457 cl->cl_d = 0;
1458 cl->cl_e = 0;
1459 cl->cl_vt = 0;
1460 cl->cl_vtadj = 0;
1461 cl->cl_cvtmin = 0;
1462 cl->cl_cvtoff = 0;
1463 cl->cl_vtperiod = 0;
1464 cl->cl_parentperiod = 0;
1465 cl->cl_f = 0;
1466 cl->cl_myf = 0;
1467 cl->cl_cfmin = 0;
1468 cl->cl_nactive = 0;
1470 cl->vt_tree = RB_ROOT;
1471 cl->cf_tree = RB_ROOT;
1472 qdisc_reset(cl->qdisc);
1474 if (cl->cl_flags & HFSC_RSC)
1475 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1476 if (cl->cl_flags & HFSC_FSC)
1477 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1478 if (cl->cl_flags & HFSC_USC)
1479 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1482 static void
1483 hfsc_reset_qdisc(struct Qdisc *sch)
1485 struct hfsc_sched *q = qdisc_priv(sch);
1486 struct hfsc_class *cl;
1487 unsigned int i;
1489 for (i = 0; i < q->clhash.hashsize; i++) {
1490 hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1491 hfsc_reset_class(cl);
1493 q->eligible = RB_ROOT;
1494 qdisc_watchdog_cancel(&q->watchdog);
1495 sch->qstats.backlog = 0;
1496 sch->q.qlen = 0;
1499 static void
1500 hfsc_destroy_qdisc(struct Qdisc *sch)
1502 struct hfsc_sched *q = qdisc_priv(sch);
1503 struct hlist_node *next;
1504 struct hfsc_class *cl;
1505 unsigned int i;
1507 for (i = 0; i < q->clhash.hashsize; i++) {
1508 hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode) {
1509 tcf_block_put(cl->block);
1510 cl->block = NULL;
1513 for (i = 0; i < q->clhash.hashsize; i++) {
1514 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1515 cl_common.hnode)
1516 hfsc_destroy_class(sch, cl);
1518 qdisc_class_hash_destroy(&q->clhash);
1519 qdisc_watchdog_cancel(&q->watchdog);
1522 static int
1523 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1525 struct hfsc_sched *q = qdisc_priv(sch);
1526 unsigned char *b = skb_tail_pointer(skb);
1527 struct tc_hfsc_qopt qopt;
1529 qopt.defcls = q->defcls;
1530 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1531 goto nla_put_failure;
1532 return skb->len;
1534 nla_put_failure:
1535 nlmsg_trim(skb, b);
1536 return -1;
1539 static int
1540 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free)
1542 struct hfsc_class *cl;
1543 int uninitialized_var(err);
1545 cl = hfsc_classify(skb, sch, &err);
1546 if (cl == NULL) {
1547 if (err & __NET_XMIT_BYPASS)
1548 qdisc_qstats_drop(sch);
1549 __qdisc_drop(skb, to_free);
1550 return err;
1553 err = qdisc_enqueue(skb, cl->qdisc, to_free);
1554 if (unlikely(err != NET_XMIT_SUCCESS)) {
1555 if (net_xmit_drop_count(err)) {
1556 cl->qstats.drops++;
1557 qdisc_qstats_drop(sch);
1559 return err;
1562 if (cl->qdisc->q.qlen == 1) {
1563 unsigned int len = qdisc_pkt_len(skb);
1565 if (cl->cl_flags & HFSC_RSC)
1566 init_ed(cl, len);
1567 if (cl->cl_flags & HFSC_FSC)
1568 init_vf(cl, len);
1570 * If this is the first packet, isolate the head so an eventual
1571 * head drop before the first dequeue operation has no chance
1572 * to invalidate the deadline.
1574 if (cl->cl_flags & HFSC_RSC)
1575 cl->qdisc->ops->peek(cl->qdisc);
1579 qdisc_qstats_backlog_inc(sch, skb);
1580 sch->q.qlen++;
1582 return NET_XMIT_SUCCESS;
1585 static struct sk_buff *
1586 hfsc_dequeue(struct Qdisc *sch)
1588 struct hfsc_sched *q = qdisc_priv(sch);
1589 struct hfsc_class *cl;
1590 struct sk_buff *skb;
1591 u64 cur_time;
1592 unsigned int next_len;
1593 int realtime = 0;
1595 if (sch->q.qlen == 0)
1596 return NULL;
1598 cur_time = psched_get_time();
1601 * if there are eligible classes, use real-time criteria.
1602 * find the class with the minimum deadline among
1603 * the eligible classes.
1605 cl = eltree_get_mindl(q, cur_time);
1606 if (cl) {
1607 realtime = 1;
1608 } else {
1610 * use link-sharing criteria
1611 * get the class with the minimum vt in the hierarchy
1613 cl = vttree_get_minvt(&q->root, cur_time);
1614 if (cl == NULL) {
1615 qdisc_qstats_overlimit(sch);
1616 hfsc_schedule_watchdog(sch);
1617 return NULL;
1621 skb = qdisc_dequeue_peeked(cl->qdisc);
1622 if (skb == NULL) {
1623 qdisc_warn_nonwc("HFSC", cl->qdisc);
1624 return NULL;
1627 bstats_update(&cl->bstats, skb);
1628 update_vf(cl, qdisc_pkt_len(skb), cur_time);
1629 if (realtime)
1630 cl->cl_cumul += qdisc_pkt_len(skb);
1632 if (cl->cl_flags & HFSC_RSC) {
1633 if (cl->qdisc->q.qlen != 0) {
1634 /* update ed */
1635 next_len = qdisc_peek_len(cl->qdisc);
1636 if (realtime)
1637 update_ed(cl, next_len);
1638 else
1639 update_d(cl, next_len);
1640 } else {
1641 /* the class becomes passive */
1642 eltree_remove(cl);
1646 qdisc_bstats_update(sch, skb);
1647 qdisc_qstats_backlog_dec(sch, skb);
1648 sch->q.qlen--;
1650 return skb;
1653 static const struct Qdisc_class_ops hfsc_class_ops = {
1654 .change = hfsc_change_class,
1655 .delete = hfsc_delete_class,
1656 .graft = hfsc_graft_class,
1657 .leaf = hfsc_class_leaf,
1658 .qlen_notify = hfsc_qlen_notify,
1659 .find = hfsc_search_class,
1660 .bind_tcf = hfsc_bind_tcf,
1661 .unbind_tcf = hfsc_unbind_tcf,
1662 .tcf_block = hfsc_tcf_block,
1663 .dump = hfsc_dump_class,
1664 .dump_stats = hfsc_dump_class_stats,
1665 .walk = hfsc_walk
1668 static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1669 .id = "hfsc",
1670 .init = hfsc_init_qdisc,
1671 .change = hfsc_change_qdisc,
1672 .reset = hfsc_reset_qdisc,
1673 .destroy = hfsc_destroy_qdisc,
1674 .dump = hfsc_dump_qdisc,
1675 .enqueue = hfsc_enqueue,
1676 .dequeue = hfsc_dequeue,
1677 .peek = qdisc_peek_dequeued,
1678 .cl_ops = &hfsc_class_ops,
1679 .priv_size = sizeof(struct hfsc_sched),
1680 .owner = THIS_MODULE
1683 static int __init
1684 hfsc_init(void)
1686 return register_qdisc(&hfsc_qdisc_ops);
1689 static void __exit
1690 hfsc_cleanup(void)
1692 unregister_qdisc(&hfsc_qdisc_ops);
1695 MODULE_LICENSE("GPL");
1696 module_init(hfsc_init);
1697 module_exit(hfsc_cleanup);