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
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.
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 */
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 {
112 struct Qdisc_class_common cl_common;
114 struct gnet_stats_basic_sync bstats;
115 struct gnet_stats_queue qstats;
116 struct net_rate_estimator __rcu *rate_est;
117 struct tcf_proto __rcu *filter_list; /* filter list */
118 struct tcf_block *block;
119 unsigned int level; /* class level in hierarchy */
121 struct hfsc_sched *sched; /* scheduler data */
122 struct hfsc_class *cl_parent; /* parent class */
123 struct list_head siblings; /* sibling classes */
124 struct list_head children; /* child classes */
125 struct Qdisc *qdisc; /* leaf qdisc */
127 struct rb_node el_node; /* qdisc's eligible tree member */
128 struct rb_root vt_tree; /* active children sorted by cl_vt */
129 struct rb_node vt_node; /* parent's vt_tree member */
130 struct rb_root cf_tree; /* active children sorted by cl_f */
131 struct rb_node cf_node; /* parent's cf_heap member */
133 u64 cl_total; /* total work in bytes */
134 u64 cl_cumul; /* cumulative work in bytes done by
135 real-time criteria */
137 u64 cl_d; /* deadline*/
138 u64 cl_e; /* eligible time */
139 u64 cl_vt; /* virtual time */
140 u64 cl_f; /* time when this class will fit for
141 link-sharing, max(myf, cfmin) */
142 u64 cl_myf; /* my fit-time (calculated from this
143 class's own upperlimit curve) */
144 u64 cl_cfmin; /* earliest children's fit-time (used
145 with cl_myf to obtain cl_f) */
146 u64 cl_cvtmin; /* minimal virtual time among the
147 children fit for link-sharing
148 (monotonic within a period) */
149 u64 cl_vtadj; /* intra-period cumulative vt
151 u64 cl_cvtoff; /* largest virtual time seen among
154 struct internal_sc cl_rsc; /* internal real-time service curve */
155 struct internal_sc cl_fsc; /* internal fair service curve */
156 struct internal_sc cl_usc; /* internal upperlimit service curve */
157 struct runtime_sc cl_deadline; /* deadline curve */
158 struct runtime_sc cl_eligible; /* eligible curve */
159 struct runtime_sc cl_virtual; /* virtual curve */
160 struct runtime_sc cl_ulimit; /* upperlimit curve */
162 u8 cl_flags; /* which curves are valid */
163 u32 cl_vtperiod; /* vt period sequence number */
164 u32 cl_parentperiod;/* parent's vt period sequence number*/
165 u32 cl_nactive; /* number of active children */
169 u16 defcls; /* default class id */
170 struct hfsc_class root; /* root class */
171 struct Qdisc_class_hash clhash; /* class hash */
172 struct rb_root eligible; /* eligible tree */
173 struct qdisc_watchdog watchdog; /* watchdog timer */
176 #define HT_INFINITY 0xffffffffffffffffULL /* infinite time value */
180 * eligible tree holds backlogged classes being sorted by their eligible times.
181 * there is one eligible tree per hfsc instance.
185 eltree_insert(struct hfsc_class *cl)
187 struct rb_node **p = &cl->sched->eligible.rb_node;
188 struct rb_node *parent = NULL;
189 struct hfsc_class *cl1;
193 cl1 = rb_entry(parent, struct hfsc_class, el_node);
194 if (cl->cl_e >= cl1->cl_e)
195 p = &parent->rb_right;
197 p = &parent->rb_left;
199 rb_link_node(&cl->el_node, parent, p);
200 rb_insert_color(&cl->el_node, &cl->sched->eligible);
204 eltree_remove(struct hfsc_class *cl)
206 rb_erase(&cl->el_node, &cl->sched->eligible);
210 eltree_update(struct hfsc_class *cl)
216 /* find the class with the minimum deadline among the eligible classes */
217 static inline struct hfsc_class *
218 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
220 struct hfsc_class *p, *cl = NULL;
223 for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
224 p = rb_entry(n, struct hfsc_class, el_node);
225 if (p->cl_e > cur_time)
227 if (cl == NULL || p->cl_d < cl->cl_d)
233 /* find the class with minimum eligible time among the eligible classes */
234 static inline struct hfsc_class *
235 eltree_get_minel(struct hfsc_sched *q)
239 n = rb_first(&q->eligible);
242 return rb_entry(n, struct hfsc_class, el_node);
246 * vttree holds holds backlogged child classes being sorted by their virtual
247 * time. each intermediate class has one vttree.
250 vttree_insert(struct hfsc_class *cl)
252 struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
253 struct rb_node *parent = NULL;
254 struct hfsc_class *cl1;
258 cl1 = rb_entry(parent, struct hfsc_class, vt_node);
259 if (cl->cl_vt >= cl1->cl_vt)
260 p = &parent->rb_right;
262 p = &parent->rb_left;
264 rb_link_node(&cl->vt_node, parent, p);
265 rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
269 vttree_remove(struct hfsc_class *cl)
271 rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
275 vttree_update(struct hfsc_class *cl)
281 static inline struct hfsc_class *
282 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
284 struct hfsc_class *p;
287 for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
288 p = rb_entry(n, struct hfsc_class, vt_node);
289 if (p->cl_f <= cur_time)
296 * get the leaf class with the minimum vt in the hierarchy
298 static struct hfsc_class *
299 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
301 /* if root-class's cfmin is bigger than cur_time nothing to do */
302 if (cl->cl_cfmin > cur_time)
305 while (cl->level > 0) {
306 cl = vttree_firstfit(cl, cur_time);
310 * update parent's cl_cvtmin.
312 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
313 cl->cl_parent->cl_cvtmin = cl->cl_vt;
319 cftree_insert(struct hfsc_class *cl)
321 struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
322 struct rb_node *parent = NULL;
323 struct hfsc_class *cl1;
327 cl1 = rb_entry(parent, struct hfsc_class, cf_node);
328 if (cl->cl_f >= cl1->cl_f)
329 p = &parent->rb_right;
331 p = &parent->rb_left;
333 rb_link_node(&cl->cf_node, parent, p);
334 rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
338 cftree_remove(struct hfsc_class *cl)
340 rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
344 cftree_update(struct hfsc_class *cl)
351 * service curve support functions
353 * external service curve parameters
356 * internal service curve parameters
357 * sm: (bytes/psched_us) << SM_SHIFT
358 * ism: (psched_us/byte) << ISM_SHIFT
361 * The clock source resolution with ktime and PSCHED_SHIFT 10 is 1.024us.
363 * sm and ism are scaled in order to keep effective digits.
364 * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
365 * digits in decimal using the following table.
367 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
368 * ------------+-------------------------------------------------------
369 * bytes/1.024us 12.8e-3 128e-3 1280e-3 12800e-3 128000e-3
371 * 1.024us/byte 78.125 7.8125 0.78125 0.078125 0.0078125
373 * So, for PSCHED_SHIFT 10 we need: SM_SHIFT 20, ISM_SHIFT 18.
375 #define SM_SHIFT (30 - PSCHED_SHIFT)
376 #define ISM_SHIFT (8 + PSCHED_SHIFT)
378 #define SM_MASK ((1ULL << SM_SHIFT) - 1)
379 #define ISM_MASK ((1ULL << ISM_SHIFT) - 1)
382 seg_x2y(u64 x, u64 sm)
388 * y = x * sm >> SM_SHIFT
389 * but divide it for the upper and lower bits to avoid overflow
391 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
396 seg_y2x(u64 y, u64 ism)
402 else if (ism == HT_INFINITY)
405 x = (y >> ISM_SHIFT) * ism
406 + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
411 /* Convert m (bps) into sm (bytes/psched us) */
417 sm = ((u64)m << SM_SHIFT);
418 sm += PSCHED_TICKS_PER_SEC - 1;
419 do_div(sm, PSCHED_TICKS_PER_SEC);
423 /* convert m (bps) into ism (psched us/byte) */
432 ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
439 /* convert d (us) into dx (psched us) */
445 dx = ((u64)d * PSCHED_TICKS_PER_SEC);
446 dx += USEC_PER_SEC - 1;
447 do_div(dx, USEC_PER_SEC);
451 /* convert sm (bytes/psched us) into m (bps) */
457 m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
461 /* convert dx (psched us) into d (us) */
467 d = dx * USEC_PER_SEC;
468 do_div(d, PSCHED_TICKS_PER_SEC);
473 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
475 isc->sm1 = m2sm(sc->m1);
476 isc->ism1 = m2ism(sc->m1);
477 isc->dx = d2dx(sc->d);
478 isc->dy = seg_x2y(isc->dx, isc->sm1);
479 isc->sm2 = m2sm(sc->m2);
480 isc->ism2 = m2ism(sc->m2);
484 * initialize the runtime service curve with the given internal
485 * service curve starting at (x, y).
488 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
492 rtsc->sm1 = isc->sm1;
493 rtsc->ism1 = isc->ism1;
496 rtsc->sm2 = isc->sm2;
497 rtsc->ism2 = isc->ism2;
501 * calculate the y-projection of the runtime service curve by the
502 * given x-projection value
505 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
511 else if (y <= rtsc->y + rtsc->dy) {
512 /* x belongs to the 1st segment */
514 x = rtsc->x + rtsc->dx;
516 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
518 /* x belongs to the 2nd segment */
519 x = rtsc->x + rtsc->dx
520 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
526 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
532 else if (x <= rtsc->x + rtsc->dx)
533 /* y belongs to the 1st segment */
534 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
536 /* y belongs to the 2nd segment */
537 y = rtsc->y + rtsc->dy
538 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
543 * update the runtime service curve by taking the minimum of the current
544 * runtime service curve and the service curve starting at (x, y).
547 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
552 if (isc->sm1 <= isc->sm2) {
553 /* service curve is convex */
554 y1 = rtsc_x2y(rtsc, x);
556 /* the current rtsc is smaller */
564 * service curve is concave
565 * compute the two y values of the current rtsc
569 y1 = rtsc_x2y(rtsc, x);
571 /* rtsc is below isc, no change to rtsc */
575 y2 = rtsc_x2y(rtsc, x + isc->dx);
576 if (y2 >= y + isc->dy) {
577 /* rtsc is above isc, replace rtsc by isc */
586 * the two curves intersect
587 * compute the offsets (dx, dy) using the reverse
588 * function of seg_x2y()
589 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
591 dx = (y1 - y) << SM_SHIFT;
592 dsm = isc->sm1 - isc->sm2;
595 * check if (x, y1) belongs to the 1st segment of rtsc.
596 * if so, add the offset.
598 if (rtsc->x + rtsc->dx > x)
599 dx += rtsc->x + rtsc->dx - x;
600 dy = seg_x2y(dx, isc->sm1);
609 init_ed(struct hfsc_class *cl, unsigned int next_len)
611 u64 cur_time = psched_get_time();
613 /* update the deadline curve */
614 rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
617 * update the eligible curve.
618 * for concave, it is equal to the deadline curve.
619 * for convex, it is a linear curve with slope m2.
621 cl->cl_eligible = cl->cl_deadline;
622 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
623 cl->cl_eligible.dx = 0;
624 cl->cl_eligible.dy = 0;
627 /* compute e and d */
628 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
629 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
635 update_ed(struct hfsc_class *cl, unsigned int next_len)
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);
644 update_d(struct hfsc_class *cl, unsigned int next_len)
646 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
650 update_cfmin(struct hfsc_class *cl)
652 struct rb_node *n = rb_first(&cl->cf_tree);
653 struct hfsc_class *p;
659 p = rb_entry(n, struct hfsc_class, cf_node);
660 cl->cl_cfmin = p->cl_f;
664 init_vf(struct hfsc_class *cl, unsigned int len)
666 struct hfsc_class *max_cl;
673 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
674 if (go_active && cl->cl_nactive++ == 0)
680 n = rb_last(&cl->cl_parent->vt_tree);
682 max_cl = rb_entry(n, struct hfsc_class, vt_node);
684 * set vt to the average of the min and max
685 * classes. if the parent's period didn't
686 * change, don't decrease vt of the class.
689 if (cl->cl_parent->cl_cvtmin != 0)
690 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
692 if (cl->cl_parent->cl_vtperiod !=
693 cl->cl_parentperiod || vt > cl->cl_vt)
697 * first child for a new parent backlog period.
698 * initialize cl_vt to the highest value seen
699 * among the siblings. this is analogous to
700 * what cur_time would provide in realtime case.
702 cl->cl_vt = cl->cl_parent->cl_cvtoff;
703 cl->cl_parent->cl_cvtmin = 0;
706 /* update the virtual curve */
707 rtsc_min(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
710 cl->cl_vtperiod++; /* increment vt period */
711 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
712 if (cl->cl_parent->cl_nactive == 0)
713 cl->cl_parentperiod++;
719 if (cl->cl_flags & HFSC_USC) {
720 /* class has upper limit curve */
722 cur_time = psched_get_time();
724 /* update the ulimit curve */
725 rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
728 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
733 f = max(cl->cl_myf, cl->cl_cfmin);
738 update_cfmin(cl->cl_parent);
743 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
745 u64 f; /* , myf_bound, delta; */
748 if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
751 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
754 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
757 if (go_passive && --cl->cl_nactive == 0)
763 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total) + cl->cl_vtadj;
766 * if vt of the class is smaller than cvtmin,
767 * the class was skipped in the past due to non-fit.
768 * if so, we need to adjust vtadj.
770 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
771 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
772 cl->cl_vt = cl->cl_parent->cl_cvtmin;
776 /* no more active child, going passive */
778 /* update cvtoff of the parent class */
779 if (cl->cl_vt > cl->cl_parent->cl_cvtoff)
780 cl->cl_parent->cl_cvtoff = cl->cl_vt;
782 /* remove this class from the vt tree */
786 update_cfmin(cl->cl_parent);
791 /* update the vt tree */
795 if (cl->cl_flags & HFSC_USC) {
796 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit, cl->cl_total);
798 cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
801 * This code causes classes to stay way under their
802 * limit when multiple classes are used at gigabit
803 * speed. needs investigation. -kaber
806 * if myf lags behind by more than one clock tick
807 * from the current time, adjust myfadj to prevent
808 * a rate-limited class from going greedy.
809 * in a steady state under rate-limiting, myf
810 * fluctuates within one clock tick.
812 myf_bound = cur_time - PSCHED_JIFFIE2US(1);
813 if (cl->cl_myf < myf_bound) {
814 delta = cur_time - cl->cl_myf;
815 cl->cl_myfadj += delta;
821 f = max(cl->cl_myf, cl->cl_cfmin);
825 update_cfmin(cl->cl_parent);
831 qdisc_peek_len(struct Qdisc *sch)
836 skb = sch->ops->peek(sch);
837 if (unlikely(skb == NULL)) {
838 qdisc_warn_nonwc("qdisc_peek_len", sch);
841 len = qdisc_pkt_len(skb);
847 hfsc_adjust_levels(struct hfsc_class *cl)
849 struct hfsc_class *p;
854 list_for_each_entry(p, &cl->children, siblings) {
855 if (p->level >= level)
856 level = p->level + 1;
859 } while ((cl = cl->cl_parent) != NULL);
862 static inline struct hfsc_class *
863 hfsc_find_class(u32 classid, struct Qdisc *sch)
865 struct hfsc_sched *q = qdisc_priv(sch);
866 struct Qdisc_class_common *clc;
868 clc = qdisc_class_find(&q->clhash, classid);
871 return container_of(clc, struct hfsc_class, cl_common);
875 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
878 sc2isc(rsc, &cl->cl_rsc);
879 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
880 cl->cl_eligible = cl->cl_deadline;
881 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
882 cl->cl_eligible.dx = 0;
883 cl->cl_eligible.dy = 0;
885 cl->cl_flags |= HFSC_RSC;
889 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
891 sc2isc(fsc, &cl->cl_fsc);
892 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
893 cl->cl_flags |= HFSC_FSC;
897 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
900 sc2isc(usc, &cl->cl_usc);
901 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
902 cl->cl_flags |= HFSC_USC;
906 hfsc_upgrade_rt(struct hfsc_class *cl)
908 cl->cl_fsc = cl->cl_rsc;
909 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
910 cl->cl_flags |= HFSC_FSC;
913 static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
914 [TCA_HFSC_RSC] = { .len = sizeof(struct tc_service_curve) },
915 [TCA_HFSC_FSC] = { .len = sizeof(struct tc_service_curve) },
916 [TCA_HFSC_USC] = { .len = sizeof(struct tc_service_curve) },
920 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
921 struct nlattr **tca, unsigned long *arg,
922 struct netlink_ext_ack *extack)
924 struct hfsc_sched *q = qdisc_priv(sch);
925 struct hfsc_class *cl = (struct hfsc_class *)*arg;
926 struct hfsc_class *parent = NULL;
927 struct nlattr *opt = tca[TCA_OPTIONS];
928 struct nlattr *tb[TCA_HFSC_MAX + 1];
929 struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
936 err = nla_parse_nested_deprecated(tb, TCA_HFSC_MAX, opt, hfsc_policy,
941 if (tb[TCA_HFSC_RSC]) {
942 rsc = nla_data(tb[TCA_HFSC_RSC]);
943 if (rsc->m1 == 0 && rsc->m2 == 0)
947 if (tb[TCA_HFSC_FSC]) {
948 fsc = nla_data(tb[TCA_HFSC_FSC]);
949 if (fsc->m1 == 0 && fsc->m2 == 0)
953 if (tb[TCA_HFSC_USC]) {
954 usc = nla_data(tb[TCA_HFSC_USC]);
955 if (usc->m1 == 0 && usc->m2 == 0)
964 cl->cl_parent->cl_common.classid != parentid)
966 if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
969 cur_time = psched_get_time();
972 err = gen_replace_estimator(&cl->bstats, NULL,
982 old_flags = cl->cl_flags;
985 hfsc_change_rsc(cl, rsc, cur_time);
987 hfsc_change_fsc(cl, fsc);
989 hfsc_change_usc(cl, usc, cur_time);
991 if (cl->qdisc->q.qlen != 0) {
992 int len = qdisc_peek_len(cl->qdisc);
994 if (cl->cl_flags & HFSC_RSC) {
995 if (old_flags & HFSC_RSC)
1001 if (cl->cl_flags & HFSC_FSC) {
1002 if (old_flags & HFSC_FSC)
1003 update_vf(cl, 0, cur_time);
1008 sch_tree_unlock(sch);
1013 if (parentid == TC_H_ROOT)
1018 parent = hfsc_find_class(parentid, sch);
1023 if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1025 if (hfsc_find_class(classid, sch))
1028 if (rsc == NULL && fsc == NULL)
1031 cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1035 err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1041 if (tca[TCA_RATE]) {
1042 err = gen_new_estimator(&cl->bstats, NULL, &cl->rate_est,
1043 NULL, true, tca[TCA_RATE]);
1045 tcf_block_put(cl->block);
1052 hfsc_change_rsc(cl, rsc, 0);
1054 hfsc_change_fsc(cl, fsc);
1056 hfsc_change_usc(cl, usc, 0);
1058 cl->cl_common.classid = classid;
1060 cl->cl_parent = parent;
1061 cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1063 if (cl->qdisc == NULL)
1064 cl->qdisc = &noop_qdisc;
1066 qdisc_hash_add(cl->qdisc, true);
1067 INIT_LIST_HEAD(&cl->children);
1068 cl->vt_tree = RB_ROOT;
1069 cl->cf_tree = RB_ROOT;
1072 /* Check if the inner class is a misconfigured 'rt' */
1073 if (!(parent->cl_flags & HFSC_FSC) && parent != &q->root) {
1074 NL_SET_ERR_MSG(extack,
1075 "Forced curve change on parent 'rt' to 'sc'");
1076 hfsc_upgrade_rt(parent);
1078 qdisc_class_hash_insert(&q->clhash, &cl->cl_common);
1079 list_add_tail(&cl->siblings, &parent->children);
1080 if (parent->level == 0)
1081 qdisc_purge_queue(parent->qdisc);
1082 hfsc_adjust_levels(parent);
1083 sch_tree_unlock(sch);
1085 qdisc_class_hash_grow(sch, &q->clhash);
1087 *arg = (unsigned long)cl;
1092 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1094 struct hfsc_sched *q = qdisc_priv(sch);
1096 tcf_block_put(cl->block);
1097 qdisc_put(cl->qdisc);
1098 gen_kill_estimator(&cl->rate_est);
1104 hfsc_delete_class(struct Qdisc *sch, unsigned long arg,
1105 struct netlink_ext_ack *extack)
1107 struct hfsc_sched *q = qdisc_priv(sch);
1108 struct hfsc_class *cl = (struct hfsc_class *)arg;
1110 if (cl->level > 0 || qdisc_class_in_use(&cl->cl_common) ||
1112 NL_SET_ERR_MSG(extack, "HFSC class in use");
1118 list_del(&cl->siblings);
1119 hfsc_adjust_levels(cl->cl_parent);
1121 qdisc_purge_queue(cl->qdisc);
1122 qdisc_class_hash_remove(&q->clhash, &cl->cl_common);
1124 sch_tree_unlock(sch);
1126 hfsc_destroy_class(sch, cl);
1130 static struct hfsc_class *
1131 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1133 struct hfsc_sched *q = qdisc_priv(sch);
1134 struct hfsc_class *head, *cl;
1135 struct tcf_result res;
1136 struct tcf_proto *tcf;
1139 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1140 (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1144 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
1146 tcf = rcu_dereference_bh(q->root.filter_list);
1147 while (tcf && (result = tcf_classify(skb, NULL, tcf, &res, false)) >= 0) {
1148 #ifdef CONFIG_NET_CLS_ACT
1153 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
1159 cl = (struct hfsc_class *)res.class;
1161 cl = hfsc_find_class(res.classid, sch);
1163 break; /* filter selected invalid classid */
1164 if (cl->level >= head->level)
1165 break; /* filter may only point downwards */
1169 return cl; /* hit leaf class */
1171 /* apply inner filter chain */
1172 tcf = rcu_dereference_bh(cl->filter_list);
1176 /* classification failed, try default class */
1177 cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1178 if (cl == NULL || cl->level > 0)
1185 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1186 struct Qdisc **old, struct netlink_ext_ack *extack)
1188 struct hfsc_class *cl = (struct hfsc_class *)arg;
1193 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1194 cl->cl_common.classid, NULL);
1199 *old = qdisc_replace(sch, new, &cl->qdisc);
1203 static struct Qdisc *
1204 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1206 struct hfsc_class *cl = (struct hfsc_class *)arg;
1215 hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1217 struct hfsc_class *cl = (struct hfsc_class *)arg;
1219 /* vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
1220 * needs to be called explicitly to remove a class from vttree.
1222 update_vf(cl, 0, 0);
1223 if (cl->cl_flags & HFSC_RSC)
1227 static unsigned long
1228 hfsc_search_class(struct Qdisc *sch, u32 classid)
1230 return (unsigned long)hfsc_find_class(classid, sch);
1233 static unsigned long
1234 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1236 struct hfsc_class *p = (struct hfsc_class *)parent;
1237 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1240 if (p != NULL && p->level <= cl->level)
1242 qdisc_class_get(&cl->cl_common);
1245 return (unsigned long)cl;
1249 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1251 struct hfsc_class *cl = (struct hfsc_class *)arg;
1253 qdisc_class_put(&cl->cl_common);
1256 static struct tcf_block *hfsc_tcf_block(struct Qdisc *sch, unsigned long arg,
1257 struct netlink_ext_ack *extack)
1259 struct hfsc_sched *q = qdisc_priv(sch);
1260 struct hfsc_class *cl = (struct hfsc_class *)arg;
1269 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1271 struct tc_service_curve tsc;
1273 tsc.m1 = sm2m(sc->sm1);
1274 tsc.d = dx2d(sc->dx);
1275 tsc.m2 = sm2m(sc->sm2);
1276 if (nla_put(skb, attr, sizeof(tsc), &tsc))
1277 goto nla_put_failure;
1286 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1288 if ((cl->cl_flags & HFSC_RSC) &&
1289 (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1290 goto nla_put_failure;
1292 if ((cl->cl_flags & HFSC_FSC) &&
1293 (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1294 goto nla_put_failure;
1296 if ((cl->cl_flags & HFSC_USC) &&
1297 (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1298 goto nla_put_failure;
1307 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1310 struct hfsc_class *cl = (struct hfsc_class *)arg;
1311 struct nlattr *nest;
1313 tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid :
1315 tcm->tcm_handle = cl->cl_common.classid;
1317 tcm->tcm_info = cl->qdisc->handle;
1319 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1321 goto nla_put_failure;
1322 if (hfsc_dump_curves(skb, cl) < 0)
1323 goto nla_put_failure;
1324 return nla_nest_end(skb, nest);
1327 nla_nest_cancel(skb, nest);
1332 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1333 struct gnet_dump *d)
1335 struct hfsc_class *cl = (struct hfsc_class *)arg;
1336 struct tc_hfsc_stats xstats;
1339 qdisc_qstats_qlen_backlog(cl->qdisc, &qlen, &cl->qstats.backlog);
1340 xstats.level = cl->level;
1341 xstats.period = cl->cl_vtperiod;
1342 xstats.work = cl->cl_total;
1343 xstats.rtwork = cl->cl_cumul;
1345 if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 ||
1346 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1347 gnet_stats_copy_queue(d, NULL, &cl->qstats, qlen) < 0)
1350 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1356 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1358 struct hfsc_sched *q = qdisc_priv(sch);
1359 struct hfsc_class *cl;
1365 for (i = 0; i < q->clhash.hashsize; i++) {
1366 hlist_for_each_entry(cl, &q->clhash.hash[i],
1368 if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg))
1375 hfsc_schedule_watchdog(struct Qdisc *sch)
1377 struct hfsc_sched *q = qdisc_priv(sch);
1378 struct hfsc_class *cl;
1381 cl = eltree_get_minel(q);
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;
1389 qdisc_watchdog_schedule(&q->watchdog, next_time);
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;
1400 qdisc_watchdog_init(&q->watchdog, sch);
1402 if (!opt || nla_len(opt) < sizeof(*qopt))
1404 qopt = nla_data(opt);
1406 q->defcls = qopt->defcls;
1407 err = qdisc_class_hash_init(&q->clhash);
1410 q->eligible = RB_ROOT;
1412 err = tcf_block_get(&q->root.block, &q->root.filter_list, sch, extack);
1416 gnet_stats_basic_sync_init(&q->root.bstats);
1417 q->root.cl_common.classid = sch->handle;
1419 q->root.qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1421 if (q->root.qdisc == NULL)
1422 q->root.qdisc = &noop_qdisc;
1424 qdisc_hash_add(q->root.qdisc, true);
1425 INIT_LIST_HEAD(&q->root.children);
1426 q->root.vt_tree = RB_ROOT;
1427 q->root.cf_tree = RB_ROOT;
1429 qdisc_class_hash_insert(&q->clhash, &q->root.cl_common);
1430 qdisc_class_hash_grow(sch, &q->clhash);
1436 hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt,
1437 struct netlink_ext_ack *extack)
1439 struct hfsc_sched *q = qdisc_priv(sch);
1440 struct tc_hfsc_qopt *qopt;
1442 if (nla_len(opt) < sizeof(*qopt))
1444 qopt = nla_data(opt);
1447 q->defcls = qopt->defcls;
1448 sch_tree_unlock(sch);
1454 hfsc_reset_class(struct hfsc_class *cl)
1464 cl->cl_vtperiod = 0;
1465 cl->cl_parentperiod = 0;
1471 cl->vt_tree = RB_ROOT;
1472 cl->cf_tree = RB_ROOT;
1473 qdisc_reset(cl->qdisc);
1475 if (cl->cl_flags & HFSC_RSC)
1476 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1477 if (cl->cl_flags & HFSC_FSC)
1478 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1479 if (cl->cl_flags & HFSC_USC)
1480 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1484 hfsc_reset_qdisc(struct Qdisc *sch)
1486 struct hfsc_sched *q = qdisc_priv(sch);
1487 struct hfsc_class *cl;
1490 for (i = 0; i < q->clhash.hashsize; i++) {
1491 hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1492 hfsc_reset_class(cl);
1494 q->eligible = RB_ROOT;
1495 qdisc_watchdog_cancel(&q->watchdog);
1499 hfsc_destroy_qdisc(struct Qdisc *sch)
1501 struct hfsc_sched *q = qdisc_priv(sch);
1502 struct hlist_node *next;
1503 struct hfsc_class *cl;
1506 for (i = 0; i < q->clhash.hashsize; i++) {
1507 hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode) {
1508 tcf_block_put(cl->block);
1512 for (i = 0; i < q->clhash.hashsize; i++) {
1513 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1515 hfsc_destroy_class(sch, cl);
1517 qdisc_class_hash_destroy(&q->clhash);
1518 qdisc_watchdog_cancel(&q->watchdog);
1522 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1524 struct hfsc_sched *q = qdisc_priv(sch);
1525 unsigned char *b = skb_tail_pointer(skb);
1526 struct tc_hfsc_qopt qopt;
1528 qopt.defcls = q->defcls;
1529 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1530 goto nla_put_failure;
1539 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free)
1541 unsigned int len = qdisc_pkt_len(skb);
1542 struct hfsc_class *cl;
1546 cl = hfsc_classify(skb, sch, &err);
1548 if (err & __NET_XMIT_BYPASS)
1549 qdisc_qstats_drop(sch);
1550 __qdisc_drop(skb, to_free);
1554 first = !cl->qdisc->q.qlen;
1555 err = qdisc_enqueue(skb, cl->qdisc, to_free);
1556 if (unlikely(err != NET_XMIT_SUCCESS)) {
1557 if (net_xmit_drop_count(err)) {
1559 qdisc_qstats_drop(sch);
1565 if (cl->cl_flags & HFSC_RSC)
1567 if (cl->cl_flags & HFSC_FSC)
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 sch->qstats.backlog += len;
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;
1592 unsigned int next_len;
1595 if (sch->q.qlen == 0)
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);
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);
1615 qdisc_qstats_overlimit(sch);
1616 hfsc_schedule_watchdog(sch);
1621 skb = qdisc_dequeue_peeked(cl->qdisc);
1623 qdisc_warn_nonwc("HFSC", cl->qdisc);
1627 bstats_update(&cl->bstats, skb);
1628 update_vf(cl, qdisc_pkt_len(skb), cur_time);
1630 cl->cl_cumul += qdisc_pkt_len(skb);
1632 if (cl->cl_flags & HFSC_RSC) {
1633 if (cl->qdisc->q.qlen != 0) {
1635 next_len = qdisc_peek_len(cl->qdisc);
1637 update_ed(cl, next_len);
1639 update_d(cl, next_len);
1641 /* the class becomes passive */
1646 qdisc_bstats_update(sch, skb);
1647 qdisc_qstats_backlog_dec(sch, 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,
1668 static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
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
1686 return register_qdisc(&hfsc_qdisc_ops);
1692 unregister_qdisc(&hfsc_qdisc_ops);
1695 MODULE_LICENSE("GPL");
1696 module_init(hfsc_init);
1697 module_exit(hfsc_cleanup);