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 qdisc_watchdog watchdog; /* watchdog timer */
190 #define HT_INFINITY 0xffffffffffffffffULL /* infinite time value */
194 * eligible tree holds backlogged classes being sorted by their eligible times.
195 * there is one eligible tree per hfsc instance.
199 eltree_insert(struct hfsc_class *cl)
201 struct rb_node **p = &cl->sched->eligible.rb_node;
202 struct rb_node *parent = NULL;
203 struct hfsc_class *cl1;
207 cl1 = rb_entry(parent, struct hfsc_class, el_node);
208 if (cl->cl_e >= cl1->cl_e)
209 p = &parent->rb_right;
211 p = &parent->rb_left;
213 rb_link_node(&cl->el_node, parent, p);
214 rb_insert_color(&cl->el_node, &cl->sched->eligible);
218 eltree_remove(struct hfsc_class *cl)
220 rb_erase(&cl->el_node, &cl->sched->eligible);
224 eltree_update(struct hfsc_class *cl)
230 /* find the class with the minimum deadline among the eligible classes */
231 static inline struct hfsc_class *
232 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
234 struct hfsc_class *p, *cl = NULL;
237 for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
238 p = rb_entry(n, struct hfsc_class, el_node);
239 if (p->cl_e > cur_time)
241 if (cl == NULL || p->cl_d < cl->cl_d)
247 /* find the class with minimum eligible time among the eligible classes */
248 static inline struct hfsc_class *
249 eltree_get_minel(struct hfsc_sched *q)
253 n = rb_first(&q->eligible);
256 return rb_entry(n, struct hfsc_class, el_node);
260 * vttree holds holds backlogged child classes being sorted by their virtual
261 * time. each intermediate class has one vttree.
264 vttree_insert(struct hfsc_class *cl)
266 struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
267 struct rb_node *parent = NULL;
268 struct hfsc_class *cl1;
272 cl1 = rb_entry(parent, struct hfsc_class, vt_node);
273 if (cl->cl_vt >= cl1->cl_vt)
274 p = &parent->rb_right;
276 p = &parent->rb_left;
278 rb_link_node(&cl->vt_node, parent, p);
279 rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
283 vttree_remove(struct hfsc_class *cl)
285 rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
289 vttree_update(struct hfsc_class *cl)
295 static inline struct hfsc_class *
296 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
298 struct hfsc_class *p;
301 for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
302 p = rb_entry(n, struct hfsc_class, vt_node);
303 if (p->cl_f <= cur_time)
310 * get the leaf class with the minimum vt in the hierarchy
312 static struct hfsc_class *
313 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
315 /* if root-class's cfmin is bigger than cur_time nothing to do */
316 if (cl->cl_cfmin > cur_time)
319 while (cl->level > 0) {
320 cl = vttree_firstfit(cl, cur_time);
324 * update parent's cl_cvtmin.
326 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
327 cl->cl_parent->cl_cvtmin = cl->cl_vt;
333 cftree_insert(struct hfsc_class *cl)
335 struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
336 struct rb_node *parent = NULL;
337 struct hfsc_class *cl1;
341 cl1 = rb_entry(parent, struct hfsc_class, cf_node);
342 if (cl->cl_f >= cl1->cl_f)
343 p = &parent->rb_right;
345 p = &parent->rb_left;
347 rb_link_node(&cl->cf_node, parent, p);
348 rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
352 cftree_remove(struct hfsc_class *cl)
354 rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
358 cftree_update(struct hfsc_class *cl)
365 * service curve support functions
367 * external service curve parameters
370 * internal service curve parameters
371 * sm: (bytes/psched_us) << SM_SHIFT
372 * ism: (psched_us/byte) << ISM_SHIFT
375 * The clock source resolution with ktime is 1.024us.
377 * sm and ism are scaled in order to keep effective digits.
378 * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
379 * digits in decimal using the following table.
381 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
382 * ------------+-------------------------------------------------------
383 * bytes/1.024us 12.8e-3 128e-3 1280e-3 12800e-3 128000e-3
385 * 1.024us/byte 78.125 7.8125 0.78125 0.078125 0.0078125
390 #define SM_MASK ((1ULL << SM_SHIFT) - 1)
391 #define ISM_MASK ((1ULL << ISM_SHIFT) - 1)
394 seg_x2y(u64 x, u64 sm)
400 * y = x * sm >> SM_SHIFT
401 * but divide it for the upper and lower bits to avoid overflow
403 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
408 seg_y2x(u64 y, u64 ism)
414 else if (ism == HT_INFINITY)
417 x = (y >> ISM_SHIFT) * ism
418 + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
423 /* Convert m (bps) into sm (bytes/psched us) */
429 sm = ((u64)m << SM_SHIFT);
430 sm += PSCHED_TICKS_PER_SEC - 1;
431 do_div(sm, PSCHED_TICKS_PER_SEC);
435 /* convert m (bps) into ism (psched us/byte) */
444 ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
451 /* convert d (us) into dx (psched us) */
457 dx = ((u64)d * PSCHED_TICKS_PER_SEC);
458 dx += USEC_PER_SEC - 1;
459 do_div(dx, USEC_PER_SEC);
463 /* convert sm (bytes/psched us) into m (bps) */
469 m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
473 /* convert dx (psched us) into d (us) */
479 d = dx * USEC_PER_SEC;
480 do_div(d, PSCHED_TICKS_PER_SEC);
485 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
487 isc->sm1 = m2sm(sc->m1);
488 isc->ism1 = m2ism(sc->m1);
489 isc->dx = d2dx(sc->d);
490 isc->dy = seg_x2y(isc->dx, isc->sm1);
491 isc->sm2 = m2sm(sc->m2);
492 isc->ism2 = m2ism(sc->m2);
496 * initialize the runtime service curve with the given internal
497 * service curve starting at (x, y).
500 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
504 rtsc->sm1 = isc->sm1;
505 rtsc->ism1 = isc->ism1;
508 rtsc->sm2 = isc->sm2;
509 rtsc->ism2 = isc->ism2;
513 * calculate the y-projection of the runtime service curve by the
514 * given x-projection value
517 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
523 else if (y <= rtsc->y + rtsc->dy) {
524 /* x belongs to the 1st segment */
526 x = rtsc->x + rtsc->dx;
528 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
530 /* x belongs to the 2nd segment */
531 x = rtsc->x + rtsc->dx
532 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
538 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
544 else if (x <= rtsc->x + rtsc->dx)
545 /* y belongs to the 1st segment */
546 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
548 /* y belongs to the 2nd segment */
549 y = rtsc->y + rtsc->dy
550 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
555 * update the runtime service curve by taking the minimum of the current
556 * runtime service curve and the service curve starting at (x, y).
559 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
564 if (isc->sm1 <= isc->sm2) {
565 /* service curve is convex */
566 y1 = rtsc_x2y(rtsc, x);
568 /* the current rtsc is smaller */
576 * service curve is concave
577 * compute the two y values of the current rtsc
581 y1 = rtsc_x2y(rtsc, x);
583 /* rtsc is below isc, no change to rtsc */
587 y2 = rtsc_x2y(rtsc, x + isc->dx);
588 if (y2 >= y + isc->dy) {
589 /* rtsc is above isc, replace rtsc by isc */
598 * the two curves intersect
599 * compute the offsets (dx, dy) using the reverse
600 * function of seg_x2y()
601 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
603 dx = (y1 - y) << SM_SHIFT;
604 dsm = isc->sm1 - isc->sm2;
607 * check if (x, y1) belongs to the 1st segment of rtsc.
608 * if so, add the offset.
610 if (rtsc->x + rtsc->dx > x)
611 dx += rtsc->x + rtsc->dx - x;
612 dy = seg_x2y(dx, isc->sm1);
622 init_ed(struct hfsc_class *cl, unsigned int next_len)
624 u64 cur_time = psched_get_time();
626 /* update the deadline curve */
627 rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
630 * update the eligible curve.
631 * for concave, it is equal to the deadline curve.
632 * for convex, it is a linear curve with slope m2.
634 cl->cl_eligible = cl->cl_deadline;
635 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
636 cl->cl_eligible.dx = 0;
637 cl->cl_eligible.dy = 0;
640 /* compute e and d */
641 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
642 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
648 update_ed(struct hfsc_class *cl, unsigned int next_len)
650 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
651 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
657 update_d(struct hfsc_class *cl, unsigned int next_len)
659 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
663 update_cfmin(struct hfsc_class *cl)
665 struct rb_node *n = rb_first(&cl->cf_tree);
666 struct hfsc_class *p;
672 p = rb_entry(n, struct hfsc_class, cf_node);
673 cl->cl_cfmin = p->cl_f;
677 init_vf(struct hfsc_class *cl, unsigned int len)
679 struct hfsc_class *max_cl;
686 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
687 if (go_active && cl->cl_nactive++ == 0)
693 n = rb_last(&cl->cl_parent->vt_tree);
695 max_cl = rb_entry(n, struct hfsc_class,vt_node);
697 * set vt to the average of the min and max
698 * classes. if the parent's period didn't
699 * change, don't decrease vt of the class.
702 if (cl->cl_parent->cl_cvtmin != 0)
703 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
705 if (cl->cl_parent->cl_vtperiod !=
706 cl->cl_parentperiod || vt > cl->cl_vt)
710 * first child for a new parent backlog period.
711 * add parent's cvtmax to cvtoff to make a new
712 * vt (vtoff + vt) larger than the vt in the
713 * last period for all children.
715 vt = cl->cl_parent->cl_cvtmax;
716 cl->cl_parent->cl_cvtoff += vt;
717 cl->cl_parent->cl_cvtmax = 0;
718 cl->cl_parent->cl_cvtmin = 0;
722 cl->cl_vtoff = cl->cl_parent->cl_cvtoff -
725 /* update the virtual curve */
726 vt = cl->cl_vt + cl->cl_vtoff;
727 rtsc_min(&cl->cl_virtual, &cl->cl_fsc, vt,
729 if (cl->cl_virtual.x == vt) {
730 cl->cl_virtual.x -= cl->cl_vtoff;
735 cl->cl_vtperiod++; /* increment vt period */
736 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
737 if (cl->cl_parent->cl_nactive == 0)
738 cl->cl_parentperiod++;
744 if (cl->cl_flags & HFSC_USC) {
745 /* class has upper limit curve */
747 cur_time = psched_get_time();
749 /* update the ulimit curve */
750 rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
753 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
759 f = max(cl->cl_myf, cl->cl_cfmin);
763 update_cfmin(cl->cl_parent);
769 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
771 u64 f; /* , myf_bound, delta; */
774 if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
777 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
780 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
783 if (go_passive && --cl->cl_nactive == 0)
789 /* no more active child, going passive */
791 /* update cvtmax of the parent class */
792 if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
793 cl->cl_parent->cl_cvtmax = cl->cl_vt;
795 /* remove this class from the vt tree */
799 update_cfmin(cl->cl_parent);
807 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
808 - cl->cl_vtoff + cl->cl_vtadj;
811 * if vt of the class is smaller than cvtmin,
812 * the class was skipped in the past due to non-fit.
813 * if so, we need to adjust vtadj.
815 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
816 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
817 cl->cl_vt = cl->cl_parent->cl_cvtmin;
820 /* update the vt tree */
823 if (cl->cl_flags & HFSC_USC) {
824 cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
828 * This code causes classes to stay way under their
829 * limit when multiple classes are used at gigabit
830 * speed. needs investigation. -kaber
833 * if myf lags behind by more than one clock tick
834 * from the current time, adjust myfadj to prevent
835 * a rate-limited class from going greedy.
836 * in a steady state under rate-limiting, myf
837 * fluctuates within one clock tick.
839 myf_bound = cur_time - PSCHED_JIFFIE2US(1);
840 if (cl->cl_myf < myf_bound) {
841 delta = cur_time - cl->cl_myf;
842 cl->cl_myfadj += delta;
848 f = max(cl->cl_myf, cl->cl_cfmin);
852 update_cfmin(cl->cl_parent);
858 set_active(struct hfsc_class *cl, unsigned int len)
860 if (cl->cl_flags & HFSC_RSC)
862 if (cl->cl_flags & HFSC_FSC)
865 list_add_tail(&cl->dlist, &cl->sched->droplist);
869 set_passive(struct hfsc_class *cl)
871 if (cl->cl_flags & HFSC_RSC)
874 list_del(&cl->dlist);
877 * vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
878 * needs to be called explicitly to remove a class from vttree.
883 qdisc_peek_len(struct Qdisc *sch)
888 skb = sch->ops->peek(sch);
890 qdisc_warn_nonwc("qdisc_peek_len", sch);
893 len = qdisc_pkt_len(skb);
899 hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl)
901 unsigned int len = cl->qdisc->q.qlen;
903 qdisc_reset(cl->qdisc);
904 qdisc_tree_decrease_qlen(cl->qdisc, len);
908 hfsc_adjust_levels(struct hfsc_class *cl)
910 struct hfsc_class *p;
915 list_for_each_entry(p, &cl->children, siblings) {
916 if (p->level >= level)
917 level = p->level + 1;
920 } while ((cl = cl->cl_parent) != NULL);
923 static inline struct hfsc_class *
924 hfsc_find_class(u32 classid, struct Qdisc *sch)
926 struct hfsc_sched *q = qdisc_priv(sch);
927 struct Qdisc_class_common *clc;
929 clc = qdisc_class_find(&q->clhash, classid);
932 return container_of(clc, struct hfsc_class, cl_common);
936 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
939 sc2isc(rsc, &cl->cl_rsc);
940 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
941 cl->cl_eligible = cl->cl_deadline;
942 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
943 cl->cl_eligible.dx = 0;
944 cl->cl_eligible.dy = 0;
946 cl->cl_flags |= HFSC_RSC;
950 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
952 sc2isc(fsc, &cl->cl_fsc);
953 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
954 cl->cl_flags |= HFSC_FSC;
958 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
961 sc2isc(usc, &cl->cl_usc);
962 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
963 cl->cl_flags |= HFSC_USC;
966 static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
967 [TCA_HFSC_RSC] = { .len = sizeof(struct tc_service_curve) },
968 [TCA_HFSC_FSC] = { .len = sizeof(struct tc_service_curve) },
969 [TCA_HFSC_USC] = { .len = sizeof(struct tc_service_curve) },
973 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
974 struct nlattr **tca, unsigned long *arg)
976 struct hfsc_sched *q = qdisc_priv(sch);
977 struct hfsc_class *cl = (struct hfsc_class *)*arg;
978 struct hfsc_class *parent = NULL;
979 struct nlattr *opt = tca[TCA_OPTIONS];
980 struct nlattr *tb[TCA_HFSC_MAX + 1];
981 struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
988 err = nla_parse_nested(tb, TCA_HFSC_MAX, opt, hfsc_policy);
992 if (tb[TCA_HFSC_RSC]) {
993 rsc = nla_data(tb[TCA_HFSC_RSC]);
994 if (rsc->m1 == 0 && rsc->m2 == 0)
998 if (tb[TCA_HFSC_FSC]) {
999 fsc = nla_data(tb[TCA_HFSC_FSC]);
1000 if (fsc->m1 == 0 && fsc->m2 == 0)
1004 if (tb[TCA_HFSC_USC]) {
1005 usc = nla_data(tb[TCA_HFSC_USC]);
1006 if (usc->m1 == 0 && usc->m2 == 0)
1012 if (cl->cl_parent &&
1013 cl->cl_parent->cl_common.classid != parentid)
1015 if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
1018 cur_time = psched_get_time();
1020 if (tca[TCA_RATE]) {
1021 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
1022 qdisc_root_sleeping_lock(sch),
1030 hfsc_change_rsc(cl, rsc, cur_time);
1032 hfsc_change_fsc(cl, fsc);
1034 hfsc_change_usc(cl, usc, cur_time);
1036 if (cl->qdisc->q.qlen != 0) {
1037 if (cl->cl_flags & HFSC_RSC)
1038 update_ed(cl, qdisc_peek_len(cl->qdisc));
1039 if (cl->cl_flags & HFSC_FSC)
1040 update_vf(cl, 0, cur_time);
1042 sch_tree_unlock(sch);
1047 if (parentid == TC_H_ROOT)
1052 parent = hfsc_find_class(parentid, sch);
1057 if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1059 if (hfsc_find_class(classid, sch))
1062 if (rsc == NULL && fsc == NULL)
1065 cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1069 if (tca[TCA_RATE]) {
1070 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
1071 qdisc_root_sleeping_lock(sch),
1080 hfsc_change_rsc(cl, rsc, 0);
1082 hfsc_change_fsc(cl, fsc);
1084 hfsc_change_usc(cl, usc, 0);
1086 cl->cl_common.classid = classid;
1089 cl->cl_parent = parent;
1090 cl->qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1091 &pfifo_qdisc_ops, classid);
1092 if (cl->qdisc == NULL)
1093 cl->qdisc = &noop_qdisc;
1094 INIT_LIST_HEAD(&cl->children);
1095 cl->vt_tree = RB_ROOT;
1096 cl->cf_tree = RB_ROOT;
1099 qdisc_class_hash_insert(&q->clhash, &cl->cl_common);
1100 list_add_tail(&cl->siblings, &parent->children);
1101 if (parent->level == 0)
1102 hfsc_purge_queue(sch, parent);
1103 hfsc_adjust_levels(parent);
1104 cl->cl_pcvtoff = parent->cl_cvtoff;
1105 sch_tree_unlock(sch);
1107 qdisc_class_hash_grow(sch, &q->clhash);
1109 *arg = (unsigned long)cl;
1114 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1116 struct hfsc_sched *q = qdisc_priv(sch);
1118 tcf_destroy_chain(&cl->filter_list);
1119 qdisc_destroy(cl->qdisc);
1120 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1126 hfsc_delete_class(struct Qdisc *sch, unsigned long arg)
1128 struct hfsc_sched *q = qdisc_priv(sch);
1129 struct hfsc_class *cl = (struct hfsc_class *)arg;
1131 if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1136 list_del(&cl->siblings);
1137 hfsc_adjust_levels(cl->cl_parent);
1139 hfsc_purge_queue(sch, cl);
1140 qdisc_class_hash_remove(&q->clhash, &cl->cl_common);
1142 BUG_ON(--cl->refcnt == 0);
1144 * This shouldn't happen: we "hold" one cops->get() when called
1145 * from tc_ctl_tclass; the destroy method is done from cops->put().
1148 sch_tree_unlock(sch);
1152 static struct hfsc_class *
1153 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1155 struct hfsc_sched *q = qdisc_priv(sch);
1156 struct hfsc_class *cl;
1157 struct tcf_result res;
1158 struct tcf_proto *tcf;
1161 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1162 (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1166 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
1167 tcf = q->root.filter_list;
1168 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
1169 #ifdef CONFIG_NET_CLS_ACT
1173 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
1178 if ((cl = (struct hfsc_class *)res.class) == NULL) {
1179 if ((cl = hfsc_find_class(res.classid, sch)) == NULL)
1180 break; /* filter selected invalid classid */
1184 return cl; /* hit leaf class */
1186 /* apply inner filter chain */
1187 tcf = cl->filter_list;
1190 /* classification failed, try default class */
1191 cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1192 if (cl == NULL || cl->level > 0)
1199 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1202 struct hfsc_class *cl = (struct hfsc_class *)arg;
1209 new = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1211 cl->cl_common.classid);
1217 hfsc_purge_queue(sch, cl);
1220 sch_tree_unlock(sch);
1224 static struct Qdisc *
1225 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1227 struct hfsc_class *cl = (struct hfsc_class *)arg;
1229 if (cl != NULL && cl->level == 0)
1236 hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1238 struct hfsc_class *cl = (struct hfsc_class *)arg;
1240 if (cl->qdisc->q.qlen == 0) {
1241 update_vf(cl, 0, 0);
1246 static unsigned long
1247 hfsc_get_class(struct Qdisc *sch, u32 classid)
1249 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1254 return (unsigned long)cl;
1258 hfsc_put_class(struct Qdisc *sch, unsigned long arg)
1260 struct hfsc_class *cl = (struct hfsc_class *)arg;
1262 if (--cl->refcnt == 0)
1263 hfsc_destroy_class(sch, cl);
1266 static unsigned long
1267 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1269 struct hfsc_class *p = (struct hfsc_class *)parent;
1270 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1273 if (p != NULL && p->level <= cl->level)
1278 return (unsigned long)cl;
1282 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1284 struct hfsc_class *cl = (struct hfsc_class *)arg;
1289 static struct tcf_proto **
1290 hfsc_tcf_chain(struct Qdisc *sch, unsigned long arg)
1292 struct hfsc_sched *q = qdisc_priv(sch);
1293 struct hfsc_class *cl = (struct hfsc_class *)arg;
1298 return &cl->filter_list;
1302 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1304 struct tc_service_curve tsc;
1306 tsc.m1 = sm2m(sc->sm1);
1307 tsc.d = dx2d(sc->dx);
1308 tsc.m2 = sm2m(sc->sm2);
1309 NLA_PUT(skb, attr, sizeof(tsc), &tsc);
1318 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1320 if ((cl->cl_flags & HFSC_RSC) &&
1321 (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1322 goto nla_put_failure;
1324 if ((cl->cl_flags & HFSC_FSC) &&
1325 (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1326 goto nla_put_failure;
1328 if ((cl->cl_flags & HFSC_USC) &&
1329 (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1330 goto nla_put_failure;
1339 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1342 struct hfsc_class *cl = (struct hfsc_class *)arg;
1343 struct nlattr *nest;
1345 tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid :
1347 tcm->tcm_handle = cl->cl_common.classid;
1349 tcm->tcm_info = cl->qdisc->handle;
1351 nest = nla_nest_start(skb, TCA_OPTIONS);
1353 goto nla_put_failure;
1354 if (hfsc_dump_curves(skb, cl) < 0)
1355 goto nla_put_failure;
1356 nla_nest_end(skb, nest);
1360 nla_nest_cancel(skb, nest);
1365 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1366 struct gnet_dump *d)
1368 struct hfsc_class *cl = (struct hfsc_class *)arg;
1369 struct tc_hfsc_stats xstats;
1371 cl->qstats.qlen = cl->qdisc->q.qlen;
1372 xstats.level = cl->level;
1373 xstats.period = cl->cl_vtperiod;
1374 xstats.work = cl->cl_total;
1375 xstats.rtwork = cl->cl_cumul;
1377 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1378 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1379 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1382 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1388 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1390 struct hfsc_sched *q = qdisc_priv(sch);
1391 struct hlist_node *n;
1392 struct hfsc_class *cl;
1398 for (i = 0; i < q->clhash.hashsize; i++) {
1399 hlist_for_each_entry(cl, n, &q->clhash.hash[i],
1401 if (arg->count < arg->skip) {
1405 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1415 hfsc_schedule_watchdog(struct Qdisc *sch)
1417 struct hfsc_sched *q = qdisc_priv(sch);
1418 struct hfsc_class *cl;
1421 if ((cl = eltree_get_minel(q)) != NULL)
1422 next_time = cl->cl_e;
1423 if (q->root.cl_cfmin != 0) {
1424 if (next_time == 0 || next_time > q->root.cl_cfmin)
1425 next_time = q->root.cl_cfmin;
1427 WARN_ON(next_time == 0);
1428 qdisc_watchdog_schedule(&q->watchdog, next_time);
1432 hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1434 struct hfsc_sched *q = qdisc_priv(sch);
1435 struct tc_hfsc_qopt *qopt;
1438 if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1440 qopt = nla_data(opt);
1442 q->defcls = qopt->defcls;
1443 err = qdisc_class_hash_init(&q->clhash);
1446 q->eligible = RB_ROOT;
1447 INIT_LIST_HEAD(&q->droplist);
1449 q->root.cl_common.classid = sch->handle;
1452 q->root.qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1455 if (q->root.qdisc == NULL)
1456 q->root.qdisc = &noop_qdisc;
1457 INIT_LIST_HEAD(&q->root.children);
1458 q->root.vt_tree = RB_ROOT;
1459 q->root.cf_tree = RB_ROOT;
1461 qdisc_class_hash_insert(&q->clhash, &q->root.cl_common);
1462 qdisc_class_hash_grow(sch, &q->clhash);
1464 qdisc_watchdog_init(&q->watchdog, sch);
1470 hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt)
1472 struct hfsc_sched *q = qdisc_priv(sch);
1473 struct tc_hfsc_qopt *qopt;
1475 if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1477 qopt = nla_data(opt);
1480 q->defcls = qopt->defcls;
1481 sch_tree_unlock(sch);
1487 hfsc_reset_class(struct hfsc_class *cl)
1500 cl->cl_vtperiod = 0;
1501 cl->cl_parentperiod = 0;
1508 cl->vt_tree = RB_ROOT;
1509 cl->cf_tree = RB_ROOT;
1510 qdisc_reset(cl->qdisc);
1512 if (cl->cl_flags & HFSC_RSC)
1513 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1514 if (cl->cl_flags & HFSC_FSC)
1515 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1516 if (cl->cl_flags & HFSC_USC)
1517 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1521 hfsc_reset_qdisc(struct Qdisc *sch)
1523 struct hfsc_sched *q = qdisc_priv(sch);
1524 struct hfsc_class *cl;
1525 struct hlist_node *n;
1528 for (i = 0; i < q->clhash.hashsize; i++) {
1529 hlist_for_each_entry(cl, n, &q->clhash.hash[i], cl_common.hnode)
1530 hfsc_reset_class(cl);
1532 q->eligible = RB_ROOT;
1533 INIT_LIST_HEAD(&q->droplist);
1534 qdisc_watchdog_cancel(&q->watchdog);
1539 hfsc_destroy_qdisc(struct Qdisc *sch)
1541 struct hfsc_sched *q = qdisc_priv(sch);
1542 struct hlist_node *n, *next;
1543 struct hfsc_class *cl;
1546 for (i = 0; i < q->clhash.hashsize; i++) {
1547 hlist_for_each_entry(cl, n, &q->clhash.hash[i], cl_common.hnode)
1548 tcf_destroy_chain(&cl->filter_list);
1550 for (i = 0; i < q->clhash.hashsize; i++) {
1551 hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[i],
1553 hfsc_destroy_class(sch, cl);
1555 qdisc_class_hash_destroy(&q->clhash);
1556 qdisc_watchdog_cancel(&q->watchdog);
1560 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1562 struct hfsc_sched *q = qdisc_priv(sch);
1563 unsigned char *b = skb_tail_pointer(skb);
1564 struct tc_hfsc_qopt qopt;
1566 qopt.defcls = q->defcls;
1567 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
1576 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1578 struct hfsc_class *cl;
1579 int uninitialized_var(err);
1581 cl = hfsc_classify(skb, sch, &err);
1583 if (err & __NET_XMIT_BYPASS)
1584 sch->qstats.drops++;
1589 err = qdisc_enqueue(skb, cl->qdisc);
1590 if (unlikely(err != NET_XMIT_SUCCESS)) {
1591 if (net_xmit_drop_count(err)) {
1593 sch->qstats.drops++;
1598 if (cl->qdisc->q.qlen == 1)
1599 set_active(cl, qdisc_pkt_len(skb));
1601 cl->bstats.packets++;
1602 cl->bstats.bytes += qdisc_pkt_len(skb);
1603 sch->bstats.packets++;
1604 sch->bstats.bytes += qdisc_pkt_len(skb);
1607 return NET_XMIT_SUCCESS;
1610 static struct sk_buff *
1611 hfsc_dequeue(struct Qdisc *sch)
1613 struct hfsc_sched *q = qdisc_priv(sch);
1614 struct hfsc_class *cl;
1615 struct sk_buff *skb;
1617 unsigned int next_len;
1620 if (sch->q.qlen == 0)
1623 cur_time = psched_get_time();
1626 * if there are eligible classes, use real-time criteria.
1627 * find the class with the minimum deadline among
1628 * the eligible classes.
1630 if ((cl = eltree_get_mindl(q, cur_time)) != NULL) {
1634 * use link-sharing criteria
1635 * get the class with the minimum vt in the hierarchy
1637 cl = vttree_get_minvt(&q->root, cur_time);
1639 sch->qstats.overlimits++;
1640 hfsc_schedule_watchdog(sch);
1645 skb = qdisc_dequeue_peeked(cl->qdisc);
1647 qdisc_warn_nonwc("HFSC", cl->qdisc);
1651 update_vf(cl, qdisc_pkt_len(skb), cur_time);
1653 cl->cl_cumul += qdisc_pkt_len(skb);
1655 if (cl->qdisc->q.qlen != 0) {
1656 if (cl->cl_flags & HFSC_RSC) {
1658 next_len = qdisc_peek_len(cl->qdisc);
1660 update_ed(cl, next_len);
1662 update_d(cl, next_len);
1665 /* the class becomes passive */
1669 sch->flags &= ~TCQ_F_THROTTLED;
1676 hfsc_drop(struct Qdisc *sch)
1678 struct hfsc_sched *q = qdisc_priv(sch);
1679 struct hfsc_class *cl;
1682 list_for_each_entry(cl, &q->droplist, dlist) {
1683 if (cl->qdisc->ops->drop != NULL &&
1684 (len = cl->qdisc->ops->drop(cl->qdisc)) > 0) {
1685 if (cl->qdisc->q.qlen == 0) {
1686 update_vf(cl, 0, 0);
1689 list_move_tail(&cl->dlist, &q->droplist);
1692 sch->qstats.drops++;
1700 static const struct Qdisc_class_ops hfsc_class_ops = {
1701 .change = hfsc_change_class,
1702 .delete = hfsc_delete_class,
1703 .graft = hfsc_graft_class,
1704 .leaf = hfsc_class_leaf,
1705 .qlen_notify = hfsc_qlen_notify,
1706 .get = hfsc_get_class,
1707 .put = hfsc_put_class,
1708 .bind_tcf = hfsc_bind_tcf,
1709 .unbind_tcf = hfsc_unbind_tcf,
1710 .tcf_chain = hfsc_tcf_chain,
1711 .dump = hfsc_dump_class,
1712 .dump_stats = hfsc_dump_class_stats,
1716 static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1718 .init = hfsc_init_qdisc,
1719 .change = hfsc_change_qdisc,
1720 .reset = hfsc_reset_qdisc,
1721 .destroy = hfsc_destroy_qdisc,
1722 .dump = hfsc_dump_qdisc,
1723 .enqueue = hfsc_enqueue,
1724 .dequeue = hfsc_dequeue,
1725 .peek = qdisc_peek_dequeued,
1727 .cl_ops = &hfsc_class_ops,
1728 .priv_size = sizeof(struct hfsc_sched),
1729 .owner = THIS_MODULE
1735 return register_qdisc(&hfsc_qdisc_ops);
1741 unregister_qdisc(&hfsc_qdisc_ops);
1744 MODULE_LICENSE("GPL");
1745 module_init(hfsc_init);
1746 module_exit(hfsc_cleanup);