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 to avoid 64-bit divide operations
81 * that are expensive on 32-bit architectures.
86 u64 sm1; /* scaled slope of the 1st segment */
87 u64 ism1; /* scaled inverse-slope of the 1st segment */
88 u64 dx; /* the x-projection of the 1st segment */
89 u64 dy; /* the y-projection of the 1st segment */
90 u64 sm2; /* scaled slope of the 2nd segment */
91 u64 ism2; /* scaled inverse-slope of the 2nd segment */
94 /* runtime service curve */
97 u64 x; /* current starting position on x-axis */
98 u64 y; /* current starting position on y-axis */
99 u64 sm1; /* scaled slope of the 1st segment */
100 u64 ism1; /* scaled inverse-slope of the 1st segment */
101 u64 dx; /* the x-projection of the 1st segment */
102 u64 dy; /* the y-projection of the 1st segment */
103 u64 sm2; /* scaled slope of the 2nd segment */
104 u64 ism2; /* scaled inverse-slope of the 2nd segment */
107 enum hfsc_class_flags
116 struct Qdisc_class_common cl_common;
117 unsigned int refcnt; /* usage count */
119 struct gnet_stats_basic bstats;
120 struct gnet_stats_queue qstats;
121 struct gnet_stats_rate_est rate_est;
122 unsigned int level; /* class level in hierarchy */
123 struct tcf_proto *filter_list; /* filter list */
124 unsigned int filter_cnt; /* filter count */
126 struct hfsc_sched *sched; /* scheduler data */
127 struct hfsc_class *cl_parent; /* parent class */
128 struct list_head siblings; /* sibling classes */
129 struct list_head children; /* child classes */
130 struct Qdisc *qdisc; /* leaf qdisc */
132 struct rb_node el_node; /* qdisc's eligible tree member */
133 struct rb_root vt_tree; /* active children sorted by cl_vt */
134 struct rb_node vt_node; /* parent's vt_tree member */
135 struct rb_root cf_tree; /* active children sorted by cl_f */
136 struct rb_node cf_node; /* parent's cf_heap member */
137 struct list_head dlist; /* drop list member */
139 u64 cl_total; /* total work in bytes */
140 u64 cl_cumul; /* cumulative work in bytes done by
141 real-time criteria */
143 u64 cl_d; /* deadline*/
144 u64 cl_e; /* eligible time */
145 u64 cl_vt; /* virtual time */
146 u64 cl_f; /* time when this class will fit for
147 link-sharing, max(myf, cfmin) */
148 u64 cl_myf; /* my fit-time (calculated from this
149 class's own upperlimit curve) */
150 u64 cl_myfadj; /* my fit-time adjustment (to cancel
151 history dependence) */
152 u64 cl_cfmin; /* earliest children's fit-time (used
153 with cl_myf to obtain cl_f) */
154 u64 cl_cvtmin; /* minimal virtual time among the
155 children fit for link-sharing
156 (monotonic within a period) */
157 u64 cl_vtadj; /* intra-period cumulative vt
159 u64 cl_vtoff; /* inter-period cumulative vt offset */
160 u64 cl_cvtmax; /* max child's vt in the last period */
161 u64 cl_cvtoff; /* cumulative cvtmax of all periods */
162 u64 cl_pcvtoff; /* parent's cvtoff at initialization
165 struct internal_sc cl_rsc; /* internal real-time service curve */
166 struct internal_sc cl_fsc; /* internal fair service curve */
167 struct internal_sc cl_usc; /* internal upperlimit service curve */
168 struct runtime_sc cl_deadline; /* deadline curve */
169 struct runtime_sc cl_eligible; /* eligible curve */
170 struct runtime_sc cl_virtual; /* virtual curve */
171 struct runtime_sc cl_ulimit; /* upperlimit curve */
173 unsigned long cl_flags; /* which curves are valid */
174 unsigned long cl_vtperiod; /* vt period sequence number */
175 unsigned long cl_parentperiod;/* parent's vt period sequence number*/
176 unsigned long cl_nactive; /* number of active children */
181 u16 defcls; /* default class id */
182 struct hfsc_class root; /* root class */
183 struct Qdisc_class_hash clhash; /* class hash */
184 struct rb_root eligible; /* eligible tree */
185 struct list_head droplist; /* active leaf class list (for
187 struct sk_buff_head requeue; /* requeued packet */
188 struct qdisc_watchdog watchdog; /* watchdog timer */
191 #define HT_INFINITY 0xffffffffffffffffULL /* infinite time value */
195 * eligible tree holds backlogged classes being sorted by their eligible times.
196 * there is one eligible tree per hfsc instance.
200 eltree_insert(struct hfsc_class *cl)
202 struct rb_node **p = &cl->sched->eligible.rb_node;
203 struct rb_node *parent = NULL;
204 struct hfsc_class *cl1;
208 cl1 = rb_entry(parent, struct hfsc_class, el_node);
209 if (cl->cl_e >= cl1->cl_e)
210 p = &parent->rb_right;
212 p = &parent->rb_left;
214 rb_link_node(&cl->el_node, parent, p);
215 rb_insert_color(&cl->el_node, &cl->sched->eligible);
219 eltree_remove(struct hfsc_class *cl)
221 rb_erase(&cl->el_node, &cl->sched->eligible);
225 eltree_update(struct hfsc_class *cl)
231 /* find the class with the minimum deadline among the eligible classes */
232 static inline struct hfsc_class *
233 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
235 struct hfsc_class *p, *cl = NULL;
238 for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
239 p = rb_entry(n, struct hfsc_class, el_node);
240 if (p->cl_e > cur_time)
242 if (cl == NULL || p->cl_d < cl->cl_d)
248 /* find the class with minimum eligible time among the eligible classes */
249 static inline struct hfsc_class *
250 eltree_get_minel(struct hfsc_sched *q)
254 n = rb_first(&q->eligible);
257 return rb_entry(n, struct hfsc_class, el_node);
261 * vttree holds holds backlogged child classes being sorted by their virtual
262 * time. each intermediate class has one vttree.
265 vttree_insert(struct hfsc_class *cl)
267 struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
268 struct rb_node *parent = NULL;
269 struct hfsc_class *cl1;
273 cl1 = rb_entry(parent, struct hfsc_class, vt_node);
274 if (cl->cl_vt >= cl1->cl_vt)
275 p = &parent->rb_right;
277 p = &parent->rb_left;
279 rb_link_node(&cl->vt_node, parent, p);
280 rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
284 vttree_remove(struct hfsc_class *cl)
286 rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
290 vttree_update(struct hfsc_class *cl)
296 static inline struct hfsc_class *
297 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
299 struct hfsc_class *p;
302 for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
303 p = rb_entry(n, struct hfsc_class, vt_node);
304 if (p->cl_f <= cur_time)
311 * get the leaf class with the minimum vt in the hierarchy
313 static struct hfsc_class *
314 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
316 /* if root-class's cfmin is bigger than cur_time nothing to do */
317 if (cl->cl_cfmin > cur_time)
320 while (cl->level > 0) {
321 cl = vttree_firstfit(cl, cur_time);
325 * update parent's cl_cvtmin.
327 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
328 cl->cl_parent->cl_cvtmin = cl->cl_vt;
334 cftree_insert(struct hfsc_class *cl)
336 struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
337 struct rb_node *parent = NULL;
338 struct hfsc_class *cl1;
342 cl1 = rb_entry(parent, struct hfsc_class, cf_node);
343 if (cl->cl_f >= cl1->cl_f)
344 p = &parent->rb_right;
346 p = &parent->rb_left;
348 rb_link_node(&cl->cf_node, parent, p);
349 rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
353 cftree_remove(struct hfsc_class *cl)
355 rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
359 cftree_update(struct hfsc_class *cl)
366 * service curve support functions
368 * external service curve parameters
371 * internal service curve parameters
372 * sm: (bytes/psched_us) << SM_SHIFT
373 * ism: (psched_us/byte) << ISM_SHIFT
376 * The clock source resolution with ktime is 1.024us.
378 * sm and ism are scaled in order to keep effective digits.
379 * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
380 * digits in decimal using the following table.
382 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
383 * ------------+-------------------------------------------------------
384 * bytes/1.024us 12.8e-3 128e-3 1280e-3 12800e-3 128000e-3
386 * 1.024us/byte 78.125 7.8125 0.78125 0.078125 0.0078125
391 #define SM_MASK ((1ULL << SM_SHIFT) - 1)
392 #define ISM_MASK ((1ULL << ISM_SHIFT) - 1)
395 seg_x2y(u64 x, u64 sm)
401 * y = x * sm >> SM_SHIFT
402 * but divide it for the upper and lower bits to avoid overflow
404 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
409 seg_y2x(u64 y, u64 ism)
415 else if (ism == HT_INFINITY)
418 x = (y >> ISM_SHIFT) * ism
419 + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
424 /* Convert m (bps) into sm (bytes/psched us) */
430 sm = ((u64)m << SM_SHIFT);
431 sm += PSCHED_TICKS_PER_SEC - 1;
432 do_div(sm, PSCHED_TICKS_PER_SEC);
436 /* convert m (bps) into ism (psched us/byte) */
445 ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
452 /* convert d (us) into dx (psched us) */
458 dx = ((u64)d * PSCHED_TICKS_PER_SEC);
459 dx += USEC_PER_SEC - 1;
460 do_div(dx, USEC_PER_SEC);
464 /* convert sm (bytes/psched us) into m (bps) */
470 m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
474 /* convert dx (psched us) into d (us) */
480 d = dx * USEC_PER_SEC;
481 do_div(d, PSCHED_TICKS_PER_SEC);
486 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
488 isc->sm1 = m2sm(sc->m1);
489 isc->ism1 = m2ism(sc->m1);
490 isc->dx = d2dx(sc->d);
491 isc->dy = seg_x2y(isc->dx, isc->sm1);
492 isc->sm2 = m2sm(sc->m2);
493 isc->ism2 = m2ism(sc->m2);
497 * initialize the runtime service curve with the given internal
498 * service curve starting at (x, y).
501 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
505 rtsc->sm1 = isc->sm1;
506 rtsc->ism1 = isc->ism1;
509 rtsc->sm2 = isc->sm2;
510 rtsc->ism2 = isc->ism2;
514 * calculate the y-projection of the runtime service curve by the
515 * given x-projection value
518 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
524 else if (y <= rtsc->y + rtsc->dy) {
525 /* x belongs to the 1st segment */
527 x = rtsc->x + rtsc->dx;
529 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
531 /* x belongs to the 2nd segment */
532 x = rtsc->x + rtsc->dx
533 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
539 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
545 else if (x <= rtsc->x + rtsc->dx)
546 /* y belongs to the 1st segment */
547 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
549 /* y belongs to the 2nd segment */
550 y = rtsc->y + rtsc->dy
551 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
556 * update the runtime service curve by taking the minimum of the current
557 * runtime service curve and the service curve starting at (x, y).
560 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
565 if (isc->sm1 <= isc->sm2) {
566 /* service curve is convex */
567 y1 = rtsc_x2y(rtsc, x);
569 /* the current rtsc is smaller */
577 * service curve is concave
578 * compute the two y values of the current rtsc
582 y1 = rtsc_x2y(rtsc, x);
584 /* rtsc is below isc, no change to rtsc */
588 y2 = rtsc_x2y(rtsc, x + isc->dx);
589 if (y2 >= y + isc->dy) {
590 /* rtsc is above isc, replace rtsc by isc */
599 * the two curves intersect
600 * compute the offsets (dx, dy) using the reverse
601 * function of seg_x2y()
602 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
604 dx = (y1 - y) << SM_SHIFT;
605 dsm = isc->sm1 - isc->sm2;
608 * check if (x, y1) belongs to the 1st segment of rtsc.
609 * if so, add the offset.
611 if (rtsc->x + rtsc->dx > x)
612 dx += rtsc->x + rtsc->dx - x;
613 dy = seg_x2y(dx, isc->sm1);
623 init_ed(struct hfsc_class *cl, unsigned int next_len)
625 u64 cur_time = psched_get_time();
627 /* update the deadline curve */
628 rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
631 * update the eligible curve.
632 * for concave, it is equal to the deadline curve.
633 * for convex, it is a linear curve with slope m2.
635 cl->cl_eligible = cl->cl_deadline;
636 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
637 cl->cl_eligible.dx = 0;
638 cl->cl_eligible.dy = 0;
641 /* compute e and d */
642 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
643 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
649 update_ed(struct hfsc_class *cl, unsigned int next_len)
651 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
652 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
658 update_d(struct hfsc_class *cl, unsigned int next_len)
660 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
664 update_cfmin(struct hfsc_class *cl)
666 struct rb_node *n = rb_first(&cl->cf_tree);
667 struct hfsc_class *p;
673 p = rb_entry(n, struct hfsc_class, cf_node);
674 cl->cl_cfmin = p->cl_f;
678 init_vf(struct hfsc_class *cl, unsigned int len)
680 struct hfsc_class *max_cl;
687 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
688 if (go_active && cl->cl_nactive++ == 0)
694 n = rb_last(&cl->cl_parent->vt_tree);
696 max_cl = rb_entry(n, struct hfsc_class,vt_node);
698 * set vt to the average of the min and max
699 * classes. if the parent's period didn't
700 * change, don't decrease vt of the class.
703 if (cl->cl_parent->cl_cvtmin != 0)
704 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
706 if (cl->cl_parent->cl_vtperiod !=
707 cl->cl_parentperiod || vt > cl->cl_vt)
711 * first child for a new parent backlog period.
712 * add parent's cvtmax to cvtoff to make a new
713 * vt (vtoff + vt) larger than the vt in the
714 * last period for all children.
716 vt = cl->cl_parent->cl_cvtmax;
717 cl->cl_parent->cl_cvtoff += vt;
718 cl->cl_parent->cl_cvtmax = 0;
719 cl->cl_parent->cl_cvtmin = 0;
723 cl->cl_vtoff = cl->cl_parent->cl_cvtoff -
726 /* update the virtual curve */
727 vt = cl->cl_vt + cl->cl_vtoff;
728 rtsc_min(&cl->cl_virtual, &cl->cl_fsc, vt,
730 if (cl->cl_virtual.x == vt) {
731 cl->cl_virtual.x -= cl->cl_vtoff;
736 cl->cl_vtperiod++; /* increment vt period */
737 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
738 if (cl->cl_parent->cl_nactive == 0)
739 cl->cl_parentperiod++;
745 if (cl->cl_flags & HFSC_USC) {
746 /* class has upper limit curve */
748 cur_time = psched_get_time();
750 /* update the ulimit curve */
751 rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
754 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
760 f = max(cl->cl_myf, cl->cl_cfmin);
764 update_cfmin(cl->cl_parent);
770 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
772 u64 f; /* , myf_bound, delta; */
775 if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
778 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
781 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
784 if (go_passive && --cl->cl_nactive == 0)
790 /* no more active child, going passive */
792 /* update cvtmax of the parent class */
793 if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
794 cl->cl_parent->cl_cvtmax = cl->cl_vt;
796 /* remove this class from the vt tree */
800 update_cfmin(cl->cl_parent);
808 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
809 - cl->cl_vtoff + cl->cl_vtadj;
812 * if vt of the class is smaller than cvtmin,
813 * the class was skipped in the past due to non-fit.
814 * if so, we need to adjust vtadj.
816 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
817 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
818 cl->cl_vt = cl->cl_parent->cl_cvtmin;
821 /* update the vt tree */
824 if (cl->cl_flags & HFSC_USC) {
825 cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
829 * This code causes classes to stay way under their
830 * limit when multiple classes are used at gigabit
831 * speed. needs investigation. -kaber
834 * if myf lags behind by more than one clock tick
835 * from the current time, adjust myfadj to prevent
836 * a rate-limited class from going greedy.
837 * in a steady state under rate-limiting, myf
838 * fluctuates within one clock tick.
840 myf_bound = cur_time - PSCHED_JIFFIE2US(1);
841 if (cl->cl_myf < myf_bound) {
842 delta = cur_time - cl->cl_myf;
843 cl->cl_myfadj += delta;
849 f = max(cl->cl_myf, cl->cl_cfmin);
853 update_cfmin(cl->cl_parent);
859 set_active(struct hfsc_class *cl, unsigned int len)
861 if (cl->cl_flags & HFSC_RSC)
863 if (cl->cl_flags & HFSC_FSC)
866 list_add_tail(&cl->dlist, &cl->sched->droplist);
870 set_passive(struct hfsc_class *cl)
872 if (cl->cl_flags & HFSC_RSC)
875 list_del(&cl->dlist);
878 * vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
879 * needs to be called explicitly to remove a class from vttree.
884 qdisc_peek_len(struct Qdisc *sch)
889 skb = sch->ops->peek(sch);
892 printk("qdisc_peek_len: non work-conserving qdisc ?\n");
895 len = qdisc_pkt_len(skb);
901 hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl)
903 unsigned int len = cl->qdisc->q.qlen;
905 qdisc_reset(cl->qdisc);
906 qdisc_tree_decrease_qlen(cl->qdisc, len);
910 hfsc_adjust_levels(struct hfsc_class *cl)
912 struct hfsc_class *p;
917 list_for_each_entry(p, &cl->children, siblings) {
918 if (p->level >= level)
919 level = p->level + 1;
922 } while ((cl = cl->cl_parent) != NULL);
925 static inline struct hfsc_class *
926 hfsc_find_class(u32 classid, struct Qdisc *sch)
928 struct hfsc_sched *q = qdisc_priv(sch);
929 struct Qdisc_class_common *clc;
931 clc = qdisc_class_find(&q->clhash, classid);
934 return container_of(clc, struct hfsc_class, cl_common);
938 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
941 sc2isc(rsc, &cl->cl_rsc);
942 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
943 cl->cl_eligible = cl->cl_deadline;
944 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
945 cl->cl_eligible.dx = 0;
946 cl->cl_eligible.dy = 0;
948 cl->cl_flags |= HFSC_RSC;
952 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
954 sc2isc(fsc, &cl->cl_fsc);
955 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
956 cl->cl_flags |= HFSC_FSC;
960 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
963 sc2isc(usc, &cl->cl_usc);
964 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
965 cl->cl_flags |= HFSC_USC;
968 static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
969 [TCA_HFSC_RSC] = { .len = sizeof(struct tc_service_curve) },
970 [TCA_HFSC_FSC] = { .len = sizeof(struct tc_service_curve) },
971 [TCA_HFSC_USC] = { .len = sizeof(struct tc_service_curve) },
975 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
976 struct nlattr **tca, unsigned long *arg)
978 struct hfsc_sched *q = qdisc_priv(sch);
979 struct hfsc_class *cl = (struct hfsc_class *)*arg;
980 struct hfsc_class *parent = NULL;
981 struct nlattr *opt = tca[TCA_OPTIONS];
982 struct nlattr *tb[TCA_HFSC_MAX + 1];
983 struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
990 err = nla_parse_nested(tb, TCA_HFSC_MAX, opt, hfsc_policy);
994 if (tb[TCA_HFSC_RSC]) {
995 rsc = nla_data(tb[TCA_HFSC_RSC]);
996 if (rsc->m1 == 0 && rsc->m2 == 0)
1000 if (tb[TCA_HFSC_FSC]) {
1001 fsc = nla_data(tb[TCA_HFSC_FSC]);
1002 if (fsc->m1 == 0 && fsc->m2 == 0)
1006 if (tb[TCA_HFSC_USC]) {
1007 usc = nla_data(tb[TCA_HFSC_USC]);
1008 if (usc->m1 == 0 && usc->m2 == 0)
1014 if (cl->cl_parent &&
1015 cl->cl_parent->cl_common.classid != parentid)
1017 if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
1020 cur_time = psched_get_time();
1024 hfsc_change_rsc(cl, rsc, cur_time);
1026 hfsc_change_fsc(cl, fsc);
1028 hfsc_change_usc(cl, usc, cur_time);
1030 if (cl->qdisc->q.qlen != 0) {
1031 if (cl->cl_flags & HFSC_RSC)
1032 update_ed(cl, qdisc_peek_len(cl->qdisc));
1033 if (cl->cl_flags & HFSC_FSC)
1034 update_vf(cl, 0, cur_time);
1036 sch_tree_unlock(sch);
1039 gen_replace_estimator(&cl->bstats, &cl->rate_est,
1040 qdisc_root_sleeping_lock(sch),
1045 if (parentid == TC_H_ROOT)
1050 parent = hfsc_find_class(parentid, sch);
1055 if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1057 if (hfsc_find_class(classid, sch))
1060 if (rsc == NULL && fsc == NULL)
1063 cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1068 hfsc_change_rsc(cl, rsc, 0);
1070 hfsc_change_fsc(cl, fsc);
1072 hfsc_change_usc(cl, usc, 0);
1074 cl->cl_common.classid = classid;
1077 cl->cl_parent = parent;
1078 cl->qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1079 &pfifo_qdisc_ops, classid);
1080 if (cl->qdisc == NULL)
1081 cl->qdisc = &noop_qdisc;
1082 INIT_LIST_HEAD(&cl->children);
1083 cl->vt_tree = RB_ROOT;
1084 cl->cf_tree = RB_ROOT;
1087 qdisc_class_hash_insert(&q->clhash, &cl->cl_common);
1088 list_add_tail(&cl->siblings, &parent->children);
1089 if (parent->level == 0)
1090 hfsc_purge_queue(sch, parent);
1091 hfsc_adjust_levels(parent);
1092 cl->cl_pcvtoff = parent->cl_cvtoff;
1093 sch_tree_unlock(sch);
1095 qdisc_class_hash_grow(sch, &q->clhash);
1098 gen_new_estimator(&cl->bstats, &cl->rate_est,
1099 qdisc_root_sleeping_lock(sch), tca[TCA_RATE]);
1100 *arg = (unsigned long)cl;
1105 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1107 struct hfsc_sched *q = qdisc_priv(sch);
1109 tcf_destroy_chain(&cl->filter_list);
1110 qdisc_destroy(cl->qdisc);
1111 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1117 hfsc_delete_class(struct Qdisc *sch, unsigned long arg)
1119 struct hfsc_sched *q = qdisc_priv(sch);
1120 struct hfsc_class *cl = (struct hfsc_class *)arg;
1122 if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1127 list_del(&cl->siblings);
1128 hfsc_adjust_levels(cl->cl_parent);
1130 hfsc_purge_queue(sch, cl);
1131 qdisc_class_hash_remove(&q->clhash, &cl->cl_common);
1133 if (--cl->refcnt == 0)
1134 hfsc_destroy_class(sch, cl);
1136 sch_tree_unlock(sch);
1140 static struct hfsc_class *
1141 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1143 struct hfsc_sched *q = qdisc_priv(sch);
1144 struct hfsc_class *cl;
1145 struct tcf_result res;
1146 struct tcf_proto *tcf;
1149 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1150 (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1154 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
1155 tcf = q->root.filter_list;
1156 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
1157 #ifdef CONFIG_NET_CLS_ACT
1161 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
1166 if ((cl = (struct hfsc_class *)res.class) == NULL) {
1167 if ((cl = hfsc_find_class(res.classid, sch)) == NULL)
1168 break; /* filter selected invalid classid */
1172 return cl; /* hit leaf class */
1174 /* apply inner filter chain */
1175 tcf = cl->filter_list;
1178 /* classification failed, try default class */
1179 cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1180 if (cl == NULL || cl->level > 0)
1187 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1190 struct hfsc_class *cl = (struct hfsc_class *)arg;
1197 new = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1199 cl->cl_common.classid);
1205 hfsc_purge_queue(sch, cl);
1206 *old = xchg(&cl->qdisc, new);
1207 sch_tree_unlock(sch);
1211 static struct Qdisc *
1212 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1214 struct hfsc_class *cl = (struct hfsc_class *)arg;
1216 if (cl != NULL && cl->level == 0)
1223 hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1225 struct hfsc_class *cl = (struct hfsc_class *)arg;
1227 if (cl->qdisc->q.qlen == 0) {
1228 update_vf(cl, 0, 0);
1233 static unsigned long
1234 hfsc_get_class(struct Qdisc *sch, u32 classid)
1236 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1241 return (unsigned long)cl;
1245 hfsc_put_class(struct Qdisc *sch, unsigned long arg)
1247 struct hfsc_class *cl = (struct hfsc_class *)arg;
1249 if (--cl->refcnt == 0)
1250 hfsc_destroy_class(sch, cl);
1253 static unsigned long
1254 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1256 struct hfsc_class *p = (struct hfsc_class *)parent;
1257 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1260 if (p != NULL && p->level <= cl->level)
1265 return (unsigned long)cl;
1269 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1271 struct hfsc_class *cl = (struct hfsc_class *)arg;
1276 static struct tcf_proto **
1277 hfsc_tcf_chain(struct Qdisc *sch, unsigned long arg)
1279 struct hfsc_sched *q = qdisc_priv(sch);
1280 struct hfsc_class *cl = (struct hfsc_class *)arg;
1285 return &cl->filter_list;
1289 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1291 struct tc_service_curve tsc;
1293 tsc.m1 = sm2m(sc->sm1);
1294 tsc.d = dx2d(sc->dx);
1295 tsc.m2 = sm2m(sc->sm2);
1296 NLA_PUT(skb, attr, sizeof(tsc), &tsc);
1305 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1307 if ((cl->cl_flags & HFSC_RSC) &&
1308 (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1309 goto nla_put_failure;
1311 if ((cl->cl_flags & HFSC_FSC) &&
1312 (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1313 goto nla_put_failure;
1315 if ((cl->cl_flags & HFSC_USC) &&
1316 (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1317 goto nla_put_failure;
1326 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1329 struct hfsc_class *cl = (struct hfsc_class *)arg;
1330 struct nlattr *nest;
1332 tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid :
1334 tcm->tcm_handle = cl->cl_common.classid;
1336 tcm->tcm_info = cl->qdisc->handle;
1338 nest = nla_nest_start(skb, TCA_OPTIONS);
1340 goto nla_put_failure;
1341 if (hfsc_dump_curves(skb, cl) < 0)
1342 goto nla_put_failure;
1343 nla_nest_end(skb, nest);
1347 nla_nest_cancel(skb, nest);
1352 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1353 struct gnet_dump *d)
1355 struct hfsc_class *cl = (struct hfsc_class *)arg;
1356 struct tc_hfsc_stats xstats;
1358 cl->qstats.qlen = cl->qdisc->q.qlen;
1359 xstats.level = cl->level;
1360 xstats.period = cl->cl_vtperiod;
1361 xstats.work = cl->cl_total;
1362 xstats.rtwork = cl->cl_cumul;
1364 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1365 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1366 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1369 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1375 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1377 struct hfsc_sched *q = qdisc_priv(sch);
1378 struct hlist_node *n;
1379 struct hfsc_class *cl;
1385 for (i = 0; i < q->clhash.hashsize; i++) {
1386 hlist_for_each_entry(cl, n, &q->clhash.hash[i],
1388 if (arg->count < arg->skip) {
1392 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1402 hfsc_schedule_watchdog(struct Qdisc *sch)
1404 struct hfsc_sched *q = qdisc_priv(sch);
1405 struct hfsc_class *cl;
1408 if ((cl = eltree_get_minel(q)) != NULL)
1409 next_time = cl->cl_e;
1410 if (q->root.cl_cfmin != 0) {
1411 if (next_time == 0 || next_time > q->root.cl_cfmin)
1412 next_time = q->root.cl_cfmin;
1414 WARN_ON(next_time == 0);
1415 qdisc_watchdog_schedule(&q->watchdog, next_time);
1419 hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1421 struct hfsc_sched *q = qdisc_priv(sch);
1422 struct tc_hfsc_qopt *qopt;
1425 if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1427 qopt = nla_data(opt);
1429 q->defcls = qopt->defcls;
1430 err = qdisc_class_hash_init(&q->clhash);
1433 q->eligible = RB_ROOT;
1434 INIT_LIST_HEAD(&q->droplist);
1435 skb_queue_head_init(&q->requeue);
1437 q->root.cl_common.classid = sch->handle;
1440 q->root.qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1443 if (q->root.qdisc == NULL)
1444 q->root.qdisc = &noop_qdisc;
1445 INIT_LIST_HEAD(&q->root.children);
1446 q->root.vt_tree = RB_ROOT;
1447 q->root.cf_tree = RB_ROOT;
1449 qdisc_class_hash_insert(&q->clhash, &q->root.cl_common);
1450 qdisc_class_hash_grow(sch, &q->clhash);
1452 qdisc_watchdog_init(&q->watchdog, sch);
1458 hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt)
1460 struct hfsc_sched *q = qdisc_priv(sch);
1461 struct tc_hfsc_qopt *qopt;
1463 if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1465 qopt = nla_data(opt);
1468 q->defcls = qopt->defcls;
1469 sch_tree_unlock(sch);
1475 hfsc_reset_class(struct hfsc_class *cl)
1488 cl->cl_vtperiod = 0;
1489 cl->cl_parentperiod = 0;
1496 cl->vt_tree = RB_ROOT;
1497 cl->cf_tree = RB_ROOT;
1498 qdisc_reset(cl->qdisc);
1500 if (cl->cl_flags & HFSC_RSC)
1501 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1502 if (cl->cl_flags & HFSC_FSC)
1503 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1504 if (cl->cl_flags & HFSC_USC)
1505 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1509 hfsc_reset_qdisc(struct Qdisc *sch)
1511 struct hfsc_sched *q = qdisc_priv(sch);
1512 struct hfsc_class *cl;
1513 struct hlist_node *n;
1516 for (i = 0; i < q->clhash.hashsize; i++) {
1517 hlist_for_each_entry(cl, n, &q->clhash.hash[i], cl_common.hnode)
1518 hfsc_reset_class(cl);
1520 __skb_queue_purge(&q->requeue);
1521 q->eligible = RB_ROOT;
1522 INIT_LIST_HEAD(&q->droplist);
1523 qdisc_watchdog_cancel(&q->watchdog);
1528 hfsc_destroy_qdisc(struct Qdisc *sch)
1530 struct hfsc_sched *q = qdisc_priv(sch);
1531 struct hlist_node *n, *next;
1532 struct hfsc_class *cl;
1535 for (i = 0; i < q->clhash.hashsize; i++) {
1536 hlist_for_each_entry(cl, n, &q->clhash.hash[i], cl_common.hnode)
1537 tcf_destroy_chain(&cl->filter_list);
1539 for (i = 0; i < q->clhash.hashsize; i++) {
1540 hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[i],
1542 hfsc_destroy_class(sch, cl);
1544 qdisc_class_hash_destroy(&q->clhash);
1545 __skb_queue_purge(&q->requeue);
1546 qdisc_watchdog_cancel(&q->watchdog);
1550 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1552 struct hfsc_sched *q = qdisc_priv(sch);
1553 unsigned char *b = skb_tail_pointer(skb);
1554 struct tc_hfsc_qopt qopt;
1556 qopt.defcls = q->defcls;
1557 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
1566 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1568 struct hfsc_class *cl;
1571 cl = hfsc_classify(skb, sch, &err);
1573 if (err & __NET_XMIT_BYPASS)
1574 sch->qstats.drops++;
1579 err = qdisc_enqueue(skb, cl->qdisc);
1580 if (unlikely(err != NET_XMIT_SUCCESS)) {
1581 if (net_xmit_drop_count(err)) {
1583 sch->qstats.drops++;
1588 if (cl->qdisc->q.qlen == 1)
1589 set_active(cl, qdisc_pkt_len(skb));
1591 cl->bstats.packets++;
1592 cl->bstats.bytes += qdisc_pkt_len(skb);
1593 sch->bstats.packets++;
1594 sch->bstats.bytes += qdisc_pkt_len(skb);
1597 return NET_XMIT_SUCCESS;
1600 static struct sk_buff *
1601 hfsc_dequeue(struct Qdisc *sch)
1603 struct hfsc_sched *q = qdisc_priv(sch);
1604 struct hfsc_class *cl;
1605 struct sk_buff *skb;
1607 unsigned int next_len;
1610 if (sch->q.qlen == 0)
1612 if ((skb = __skb_dequeue(&q->requeue)))
1615 cur_time = psched_get_time();
1618 * if there are eligible classes, use real-time criteria.
1619 * find the class with the minimum deadline among
1620 * the eligible classes.
1622 if ((cl = eltree_get_mindl(q, cur_time)) != NULL) {
1626 * use link-sharing criteria
1627 * get the class with the minimum vt in the hierarchy
1629 cl = vttree_get_minvt(&q->root, cur_time);
1631 sch->qstats.overlimits++;
1632 hfsc_schedule_watchdog(sch);
1637 skb = qdisc_dequeue_peeked(cl->qdisc);
1639 if (net_ratelimit())
1640 printk("HFSC: Non-work-conserving qdisc ?\n");
1644 update_vf(cl, qdisc_pkt_len(skb), cur_time);
1646 cl->cl_cumul += qdisc_pkt_len(skb);
1648 if (cl->qdisc->q.qlen != 0) {
1649 if (cl->cl_flags & HFSC_RSC) {
1651 next_len = qdisc_peek_len(cl->qdisc);
1653 update_ed(cl, next_len);
1655 update_d(cl, next_len);
1658 /* the class becomes passive */
1663 sch->flags &= ~TCQ_F_THROTTLED;
1670 hfsc_requeue(struct sk_buff *skb, struct Qdisc *sch)
1672 struct hfsc_sched *q = qdisc_priv(sch);
1674 __skb_queue_head(&q->requeue, skb);
1676 sch->qstats.requeues++;
1677 return NET_XMIT_SUCCESS;
1681 hfsc_drop(struct Qdisc *sch)
1683 struct hfsc_sched *q = qdisc_priv(sch);
1684 struct hfsc_class *cl;
1687 list_for_each_entry(cl, &q->droplist, dlist) {
1688 if (cl->qdisc->ops->drop != NULL &&
1689 (len = cl->qdisc->ops->drop(cl->qdisc)) > 0) {
1690 if (cl->qdisc->q.qlen == 0) {
1691 update_vf(cl, 0, 0);
1694 list_move_tail(&cl->dlist, &q->droplist);
1697 sch->qstats.drops++;
1705 static const struct Qdisc_class_ops hfsc_class_ops = {
1706 .change = hfsc_change_class,
1707 .delete = hfsc_delete_class,
1708 .graft = hfsc_graft_class,
1709 .leaf = hfsc_class_leaf,
1710 .qlen_notify = hfsc_qlen_notify,
1711 .get = hfsc_get_class,
1712 .put = hfsc_put_class,
1713 .bind_tcf = hfsc_bind_tcf,
1714 .unbind_tcf = hfsc_unbind_tcf,
1715 .tcf_chain = hfsc_tcf_chain,
1716 .dump = hfsc_dump_class,
1717 .dump_stats = hfsc_dump_class_stats,
1721 static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1723 .init = hfsc_init_qdisc,
1724 .change = hfsc_change_qdisc,
1725 .reset = hfsc_reset_qdisc,
1726 .destroy = hfsc_destroy_qdisc,
1727 .dump = hfsc_dump_qdisc,
1728 .enqueue = hfsc_enqueue,
1729 .dequeue = hfsc_dequeue,
1730 .peek = qdisc_peek_dequeued,
1731 .requeue = hfsc_requeue,
1733 .cl_ops = &hfsc_class_ops,
1734 .priv_size = sizeof(struct hfsc_sched),
1735 .owner = THIS_MODULE
1741 return register_qdisc(&hfsc_qdisc_ops);
1747 unregister_qdisc(&hfsc_qdisc_ops);
1750 MODULE_LICENSE("GPL");
1751 module_init(hfsc_init);
1752 module_exit(hfsc_cleanup);