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;
113 unsigned int refcnt; /* usage count */
115 struct gnet_stats_basic_packed bstats;
116 struct gnet_stats_queue qstats;
117 struct gnet_stats_rate_est64 rate_est;
118 struct tcf_proto __rcu *filter_list; /* filter list */
119 unsigned int filter_cnt; /* filter count */
120 unsigned int level; /* class level in hierarchy */
122 struct hfsc_sched *sched; /* scheduler data */
123 struct hfsc_class *cl_parent; /* parent class */
124 struct list_head siblings; /* sibling classes */
125 struct list_head children; /* child classes */
126 struct Qdisc *qdisc; /* leaf qdisc */
128 struct rb_node el_node; /* qdisc's eligible tree member */
129 struct rb_root vt_tree; /* active children sorted by cl_vt */
130 struct rb_node vt_node; /* parent's vt_tree member */
131 struct rb_root cf_tree; /* active children sorted by cl_f */
132 struct rb_node cf_node; /* parent's cf_heap member */
134 u64 cl_total; /* total work in bytes */
135 u64 cl_cumul; /* cumulative work in bytes done by
136 real-time criteria */
138 u64 cl_d; /* deadline*/
139 u64 cl_e; /* eligible time */
140 u64 cl_vt; /* virtual time */
141 u64 cl_f; /* time when this class will fit for
142 link-sharing, max(myf, cfmin) */
143 u64 cl_myf; /* my fit-time (calculated from this
144 class's own upperlimit curve) */
145 u64 cl_cfmin; /* earliest children's fit-time (used
146 with cl_myf to obtain cl_f) */
147 u64 cl_cvtmin; /* minimal virtual time among the
148 children fit for link-sharing
149 (monotonic within a period) */
150 u64 cl_vtadj; /* intra-period cumulative vt
152 u64 cl_cvtoff; /* largest virtual time seen among
155 struct internal_sc cl_rsc; /* internal real-time service curve */
156 struct internal_sc cl_fsc; /* internal fair service curve */
157 struct internal_sc cl_usc; /* internal upperlimit service curve */
158 struct runtime_sc cl_deadline; /* deadline curve */
159 struct runtime_sc cl_eligible; /* eligible curve */
160 struct runtime_sc cl_virtual; /* virtual curve */
161 struct runtime_sc cl_ulimit; /* upperlimit curve */
163 u8 cl_flags; /* which curves are valid */
164 u32 cl_vtperiod; /* vt period sequence number */
165 u32 cl_parentperiod;/* parent's vt period sequence number*/
166 u32 cl_nactive; /* number of active children */
170 u16 defcls; /* default class id */
171 struct hfsc_class root; /* root class */
172 struct Qdisc_class_hash clhash; /* class hash */
173 struct rb_root eligible; /* eligible tree */
174 struct qdisc_watchdog watchdog; /* watchdog timer */
177 #define HT_INFINITY 0xffffffffffffffffULL /* infinite time value */
181 * eligible tree holds backlogged classes being sorted by their eligible times.
182 * there is one eligible tree per hfsc instance.
186 eltree_insert(struct hfsc_class *cl)
188 struct rb_node **p = &cl->sched->eligible.rb_node;
189 struct rb_node *parent = NULL;
190 struct hfsc_class *cl1;
194 cl1 = rb_entry(parent, struct hfsc_class, el_node);
195 if (cl->cl_e >= cl1->cl_e)
196 p = &parent->rb_right;
198 p = &parent->rb_left;
200 rb_link_node(&cl->el_node, parent, p);
201 rb_insert_color(&cl->el_node, &cl->sched->eligible);
205 eltree_remove(struct hfsc_class *cl)
207 rb_erase(&cl->el_node, &cl->sched->eligible);
211 eltree_update(struct hfsc_class *cl)
217 /* find the class with the minimum deadline among the eligible classes */
218 static inline struct hfsc_class *
219 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
221 struct hfsc_class *p, *cl = NULL;
224 for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
225 p = rb_entry(n, struct hfsc_class, el_node);
226 if (p->cl_e > cur_time)
228 if (cl == NULL || p->cl_d < cl->cl_d)
234 /* find the class with minimum eligible time among the eligible classes */
235 static inline struct hfsc_class *
236 eltree_get_minel(struct hfsc_sched *q)
240 n = rb_first(&q->eligible);
243 return rb_entry(n, struct hfsc_class, el_node);
247 * vttree holds holds backlogged child classes being sorted by their virtual
248 * time. each intermediate class has one vttree.
251 vttree_insert(struct hfsc_class *cl)
253 struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
254 struct rb_node *parent = NULL;
255 struct hfsc_class *cl1;
259 cl1 = rb_entry(parent, struct hfsc_class, vt_node);
260 if (cl->cl_vt >= cl1->cl_vt)
261 p = &parent->rb_right;
263 p = &parent->rb_left;
265 rb_link_node(&cl->vt_node, parent, p);
266 rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
270 vttree_remove(struct hfsc_class *cl)
272 rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
276 vttree_update(struct hfsc_class *cl)
282 static inline struct hfsc_class *
283 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
285 struct hfsc_class *p;
288 for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
289 p = rb_entry(n, struct hfsc_class, vt_node);
290 if (p->cl_f <= cur_time)
297 * get the leaf class with the minimum vt in the hierarchy
299 static struct hfsc_class *
300 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
302 /* if root-class's cfmin is bigger than cur_time nothing to do */
303 if (cl->cl_cfmin > cur_time)
306 while (cl->level > 0) {
307 cl = vttree_firstfit(cl, cur_time);
311 * update parent's cl_cvtmin.
313 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
314 cl->cl_parent->cl_cvtmin = cl->cl_vt;
320 cftree_insert(struct hfsc_class *cl)
322 struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
323 struct rb_node *parent = NULL;
324 struct hfsc_class *cl1;
328 cl1 = rb_entry(parent, struct hfsc_class, cf_node);
329 if (cl->cl_f >= cl1->cl_f)
330 p = &parent->rb_right;
332 p = &parent->rb_left;
334 rb_link_node(&cl->cf_node, parent, p);
335 rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
339 cftree_remove(struct hfsc_class *cl)
341 rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
345 cftree_update(struct hfsc_class *cl)
352 * service curve support functions
354 * external service curve parameters
357 * internal service curve parameters
358 * sm: (bytes/psched_us) << SM_SHIFT
359 * ism: (psched_us/byte) << ISM_SHIFT
362 * The clock source resolution with ktime and PSCHED_SHIFT 10 is 1.024us.
364 * sm and ism are scaled in order to keep effective digits.
365 * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
366 * digits in decimal using the following table.
368 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
369 * ------------+-------------------------------------------------------
370 * bytes/1.024us 12.8e-3 128e-3 1280e-3 12800e-3 128000e-3
372 * 1.024us/byte 78.125 7.8125 0.78125 0.078125 0.0078125
374 * So, for PSCHED_SHIFT 10 we need: SM_SHIFT 20, ISM_SHIFT 18.
376 #define SM_SHIFT (30 - PSCHED_SHIFT)
377 #define ISM_SHIFT (8 + PSCHED_SHIFT)
379 #define SM_MASK ((1ULL << SM_SHIFT) - 1)
380 #define ISM_MASK ((1ULL << ISM_SHIFT) - 1)
383 seg_x2y(u64 x, u64 sm)
389 * y = x * sm >> SM_SHIFT
390 * but divide it for the upper and lower bits to avoid overflow
392 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
397 seg_y2x(u64 y, u64 ism)
403 else if (ism == HT_INFINITY)
406 x = (y >> ISM_SHIFT) * ism
407 + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
412 /* Convert m (bps) into sm (bytes/psched us) */
418 sm = ((u64)m << SM_SHIFT);
419 sm += PSCHED_TICKS_PER_SEC - 1;
420 do_div(sm, PSCHED_TICKS_PER_SEC);
424 /* convert m (bps) into ism (psched us/byte) */
433 ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
440 /* convert d (us) into dx (psched us) */
446 dx = ((u64)d * PSCHED_TICKS_PER_SEC);
447 dx += USEC_PER_SEC - 1;
448 do_div(dx, USEC_PER_SEC);
452 /* convert sm (bytes/psched us) into m (bps) */
458 m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
462 /* convert dx (psched us) into d (us) */
468 d = dx * USEC_PER_SEC;
469 do_div(d, PSCHED_TICKS_PER_SEC);
474 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
476 isc->sm1 = m2sm(sc->m1);
477 isc->ism1 = m2ism(sc->m1);
478 isc->dx = d2dx(sc->d);
479 isc->dy = seg_x2y(isc->dx, isc->sm1);
480 isc->sm2 = m2sm(sc->m2);
481 isc->ism2 = m2ism(sc->m2);
485 * initialize the runtime service curve with the given internal
486 * service curve starting at (x, y).
489 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
493 rtsc->sm1 = isc->sm1;
494 rtsc->ism1 = isc->ism1;
497 rtsc->sm2 = isc->sm2;
498 rtsc->ism2 = isc->ism2;
502 * calculate the y-projection of the runtime service curve by the
503 * given x-projection value
506 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
512 else if (y <= rtsc->y + rtsc->dy) {
513 /* x belongs to the 1st segment */
515 x = rtsc->x + rtsc->dx;
517 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
519 /* x belongs to the 2nd segment */
520 x = rtsc->x + rtsc->dx
521 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
527 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
533 else if (x <= rtsc->x + rtsc->dx)
534 /* y belongs to the 1st segment */
535 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
537 /* y belongs to the 2nd segment */
538 y = rtsc->y + rtsc->dy
539 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
544 * update the runtime service curve by taking the minimum of the current
545 * runtime service curve and the service curve starting at (x, y).
548 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
553 if (isc->sm1 <= isc->sm2) {
554 /* service curve is convex */
555 y1 = rtsc_x2y(rtsc, x);
557 /* the current rtsc is smaller */
565 * service curve is concave
566 * compute the two y values of the current rtsc
570 y1 = rtsc_x2y(rtsc, x);
572 /* rtsc is below isc, no change to rtsc */
576 y2 = rtsc_x2y(rtsc, x + isc->dx);
577 if (y2 >= y + isc->dy) {
578 /* rtsc is above isc, replace rtsc by isc */
587 * the two curves intersect
588 * compute the offsets (dx, dy) using the reverse
589 * function of seg_x2y()
590 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
592 dx = (y1 - y) << SM_SHIFT;
593 dsm = isc->sm1 - isc->sm2;
596 * check if (x, y1) belongs to the 1st segment of rtsc.
597 * if so, add the offset.
599 if (rtsc->x + rtsc->dx > x)
600 dx += rtsc->x + rtsc->dx - x;
601 dy = seg_x2y(dx, isc->sm1);
610 init_ed(struct hfsc_class *cl, unsigned int next_len)
612 u64 cur_time = psched_get_time();
614 /* update the deadline curve */
615 rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
618 * update the eligible curve.
619 * for concave, it is equal to the deadline curve.
620 * for convex, it is a linear curve with slope m2.
622 cl->cl_eligible = cl->cl_deadline;
623 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
624 cl->cl_eligible.dx = 0;
625 cl->cl_eligible.dy = 0;
628 /* compute e and d */
629 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
630 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
636 update_ed(struct hfsc_class *cl, unsigned int next_len)
638 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
639 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
645 update_d(struct hfsc_class *cl, unsigned int next_len)
647 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
651 update_cfmin(struct hfsc_class *cl)
653 struct rb_node *n = rb_first(&cl->cf_tree);
654 struct hfsc_class *p;
660 p = rb_entry(n, struct hfsc_class, cf_node);
661 cl->cl_cfmin = p->cl_f;
665 init_vf(struct hfsc_class *cl, unsigned int len)
667 struct hfsc_class *max_cl;
674 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
675 if (go_active && cl->cl_nactive++ == 0)
681 n = rb_last(&cl->cl_parent->vt_tree);
683 max_cl = rb_entry(n, struct hfsc_class, vt_node);
685 * set vt to the average of the min and max
686 * classes. if the parent's period didn't
687 * change, don't decrease vt of the class.
690 if (cl->cl_parent->cl_cvtmin != 0)
691 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
693 if (cl->cl_parent->cl_vtperiod !=
694 cl->cl_parentperiod || vt > cl->cl_vt)
698 * first child for a new parent backlog period.
699 * initialize cl_vt to the highest value seen
700 * among the siblings. this is analogous to
701 * what cur_time would provide in realtime case.
703 cl->cl_vt = cl->cl_parent->cl_cvtoff;
704 cl->cl_parent->cl_cvtmin = 0;
707 /* update the virtual curve */
708 rtsc_min(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
711 cl->cl_vtperiod++; /* increment vt period */
712 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
713 if (cl->cl_parent->cl_nactive == 0)
714 cl->cl_parentperiod++;
720 if (cl->cl_flags & HFSC_USC) {
721 /* class has upper limit curve */
723 cur_time = psched_get_time();
725 /* update the ulimit curve */
726 rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
729 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
734 f = max(cl->cl_myf, cl->cl_cfmin);
739 update_cfmin(cl->cl_parent);
744 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
746 u64 f; /* , myf_bound, delta; */
749 if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
752 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
755 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
758 if (go_passive && --cl->cl_nactive == 0)
764 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total) + cl->cl_vtadj;
767 * if vt of the class is smaller than cvtmin,
768 * the class was skipped in the past due to non-fit.
769 * if so, we need to adjust vtadj.
771 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
772 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
773 cl->cl_vt = cl->cl_parent->cl_cvtmin;
777 /* no more active child, going passive */
779 /* update cvtoff of the parent class */
780 if (cl->cl_vt > cl->cl_parent->cl_cvtoff)
781 cl->cl_parent->cl_cvtoff = cl->cl_vt;
783 /* remove this class from the vt tree */
787 update_cfmin(cl->cl_parent);
792 /* update the vt tree */
796 if (cl->cl_flags & HFSC_USC) {
797 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit, cl->cl_total);
799 cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
802 * This code causes classes to stay way under their
803 * limit when multiple classes are used at gigabit
804 * speed. needs investigation. -kaber
807 * if myf lags behind by more than one clock tick
808 * from the current time, adjust myfadj to prevent
809 * a rate-limited class from going greedy.
810 * in a steady state under rate-limiting, myf
811 * fluctuates within one clock tick.
813 myf_bound = cur_time - PSCHED_JIFFIE2US(1);
814 if (cl->cl_myf < myf_bound) {
815 delta = cur_time - cl->cl_myf;
816 cl->cl_myfadj += delta;
822 f = max(cl->cl_myf, cl->cl_cfmin);
826 update_cfmin(cl->cl_parent);
832 set_active(struct hfsc_class *cl, unsigned int len)
834 if (cl->cl_flags & HFSC_RSC)
836 if (cl->cl_flags & HFSC_FSC)
842 set_passive(struct hfsc_class *cl)
844 if (cl->cl_flags & HFSC_RSC)
848 * vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
849 * needs to be called explicitly to remove a class from vttree.
854 qdisc_peek_len(struct Qdisc *sch)
859 skb = sch->ops->peek(sch);
860 if (unlikely(skb == NULL)) {
861 qdisc_warn_nonwc("qdisc_peek_len", sch);
864 len = qdisc_pkt_len(skb);
870 hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl)
872 unsigned int len = cl->qdisc->q.qlen;
873 unsigned int backlog = cl->qdisc->qstats.backlog;
875 qdisc_reset(cl->qdisc);
876 qdisc_tree_reduce_backlog(cl->qdisc, len, backlog);
880 hfsc_adjust_levels(struct hfsc_class *cl)
882 struct hfsc_class *p;
887 list_for_each_entry(p, &cl->children, siblings) {
888 if (p->level >= level)
889 level = p->level + 1;
892 } while ((cl = cl->cl_parent) != NULL);
895 static inline struct hfsc_class *
896 hfsc_find_class(u32 classid, struct Qdisc *sch)
898 struct hfsc_sched *q = qdisc_priv(sch);
899 struct Qdisc_class_common *clc;
901 clc = qdisc_class_find(&q->clhash, classid);
904 return container_of(clc, struct hfsc_class, cl_common);
908 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
911 sc2isc(rsc, &cl->cl_rsc);
912 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
913 cl->cl_eligible = cl->cl_deadline;
914 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
915 cl->cl_eligible.dx = 0;
916 cl->cl_eligible.dy = 0;
918 cl->cl_flags |= HFSC_RSC;
922 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
924 sc2isc(fsc, &cl->cl_fsc);
925 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
926 cl->cl_flags |= HFSC_FSC;
930 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
933 sc2isc(usc, &cl->cl_usc);
934 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
935 cl->cl_flags |= HFSC_USC;
938 static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
939 [TCA_HFSC_RSC] = { .len = sizeof(struct tc_service_curve) },
940 [TCA_HFSC_FSC] = { .len = sizeof(struct tc_service_curve) },
941 [TCA_HFSC_USC] = { .len = sizeof(struct tc_service_curve) },
945 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
946 struct nlattr **tca, unsigned long *arg)
948 struct hfsc_sched *q = qdisc_priv(sch);
949 struct hfsc_class *cl = (struct hfsc_class *)*arg;
950 struct hfsc_class *parent = NULL;
951 struct nlattr *opt = tca[TCA_OPTIONS];
952 struct nlattr *tb[TCA_HFSC_MAX + 1];
953 struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
960 err = nla_parse_nested(tb, TCA_HFSC_MAX, opt, hfsc_policy);
964 if (tb[TCA_HFSC_RSC]) {
965 rsc = nla_data(tb[TCA_HFSC_RSC]);
966 if (rsc->m1 == 0 && rsc->m2 == 0)
970 if (tb[TCA_HFSC_FSC]) {
971 fsc = nla_data(tb[TCA_HFSC_FSC]);
972 if (fsc->m1 == 0 && fsc->m2 == 0)
976 if (tb[TCA_HFSC_USC]) {
977 usc = nla_data(tb[TCA_HFSC_USC]);
978 if (usc->m1 == 0 && usc->m2 == 0)
985 cl->cl_parent->cl_common.classid != parentid)
987 if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
990 cur_time = psched_get_time();
993 err = gen_replace_estimator(&cl->bstats, NULL,
996 qdisc_root_sleeping_running(sch),
1004 hfsc_change_rsc(cl, rsc, cur_time);
1006 hfsc_change_fsc(cl, fsc);
1008 hfsc_change_usc(cl, usc, cur_time);
1010 if (cl->qdisc->q.qlen != 0) {
1011 if (cl->cl_flags & HFSC_RSC)
1012 update_ed(cl, qdisc_peek_len(cl->qdisc));
1013 if (cl->cl_flags & HFSC_FSC)
1014 update_vf(cl, 0, cur_time);
1016 sch_tree_unlock(sch);
1021 if (parentid == TC_H_ROOT)
1026 parent = hfsc_find_class(parentid, sch);
1031 if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1033 if (hfsc_find_class(classid, sch))
1036 if (rsc == NULL && fsc == NULL)
1039 cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1043 if (tca[TCA_RATE]) {
1044 err = gen_new_estimator(&cl->bstats, NULL, &cl->rate_est,
1046 qdisc_root_sleeping_running(sch),
1055 hfsc_change_rsc(cl, rsc, 0);
1057 hfsc_change_fsc(cl, fsc);
1059 hfsc_change_usc(cl, usc, 0);
1061 cl->cl_common.classid = classid;
1064 cl->cl_parent = parent;
1065 cl->qdisc = qdisc_create_dflt(sch->dev_queue,
1066 &pfifo_qdisc_ops, classid);
1067 if (cl->qdisc == NULL)
1068 cl->qdisc = &noop_qdisc;
1069 INIT_LIST_HEAD(&cl->children);
1070 cl->vt_tree = RB_ROOT;
1071 cl->cf_tree = RB_ROOT;
1074 qdisc_class_hash_insert(&q->clhash, &cl->cl_common);
1075 list_add_tail(&cl->siblings, &parent->children);
1076 if (parent->level == 0)
1077 hfsc_purge_queue(sch, parent);
1078 hfsc_adjust_levels(parent);
1079 sch_tree_unlock(sch);
1081 qdisc_class_hash_grow(sch, &q->clhash);
1083 *arg = (unsigned long)cl;
1088 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1090 struct hfsc_sched *q = qdisc_priv(sch);
1092 tcf_destroy_chain(&cl->filter_list);
1093 qdisc_destroy(cl->qdisc);
1094 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1100 hfsc_delete_class(struct Qdisc *sch, unsigned long arg)
1102 struct hfsc_sched *q = qdisc_priv(sch);
1103 struct hfsc_class *cl = (struct hfsc_class *)arg;
1105 if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1110 list_del(&cl->siblings);
1111 hfsc_adjust_levels(cl->cl_parent);
1113 hfsc_purge_queue(sch, cl);
1114 qdisc_class_hash_remove(&q->clhash, &cl->cl_common);
1116 BUG_ON(--cl->refcnt == 0);
1118 * This shouldn't happen: we "hold" one cops->get() when called
1119 * from tc_ctl_tclass; the destroy method is done from cops->put().
1122 sch_tree_unlock(sch);
1126 static struct hfsc_class *
1127 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1129 struct hfsc_sched *q = qdisc_priv(sch);
1130 struct hfsc_class *head, *cl;
1131 struct tcf_result res;
1132 struct tcf_proto *tcf;
1135 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1136 (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1140 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
1142 tcf = rcu_dereference_bh(q->root.filter_list);
1143 while (tcf && (result = tc_classify(skb, tcf, &res, false)) >= 0) {
1144 #ifdef CONFIG_NET_CLS_ACT
1148 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
1153 cl = (struct hfsc_class *)res.class;
1155 cl = hfsc_find_class(res.classid, sch);
1157 break; /* filter selected invalid classid */
1158 if (cl->level >= head->level)
1159 break; /* filter may only point downwards */
1163 return cl; /* hit leaf class */
1165 /* apply inner filter chain */
1166 tcf = rcu_dereference_bh(cl->filter_list);
1170 /* classification failed, try default class */
1171 cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1172 if (cl == NULL || cl->level > 0)
1179 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1182 struct hfsc_class *cl = (struct hfsc_class *)arg;
1187 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1188 cl->cl_common.classid);
1193 *old = qdisc_replace(sch, new, &cl->qdisc);
1197 static struct Qdisc *
1198 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1200 struct hfsc_class *cl = (struct hfsc_class *)arg;
1209 hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1211 struct hfsc_class *cl = (struct hfsc_class *)arg;
1213 if (cl->qdisc->q.qlen == 0) {
1214 update_vf(cl, 0, 0);
1219 static unsigned long
1220 hfsc_get_class(struct Qdisc *sch, u32 classid)
1222 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1227 return (unsigned long)cl;
1231 hfsc_put_class(struct Qdisc *sch, unsigned long arg)
1233 struct hfsc_class *cl = (struct hfsc_class *)arg;
1235 if (--cl->refcnt == 0)
1236 hfsc_destroy_class(sch, cl);
1239 static unsigned long
1240 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1242 struct hfsc_class *p = (struct hfsc_class *)parent;
1243 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1246 if (p != NULL && p->level <= cl->level)
1251 return (unsigned long)cl;
1255 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1257 struct hfsc_class *cl = (struct hfsc_class *)arg;
1262 static struct tcf_proto __rcu **
1263 hfsc_tcf_chain(struct Qdisc *sch, unsigned long arg)
1265 struct hfsc_sched *q = qdisc_priv(sch);
1266 struct hfsc_class *cl = (struct hfsc_class *)arg;
1271 return &cl->filter_list;
1275 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1277 struct tc_service_curve tsc;
1279 tsc.m1 = sm2m(sc->sm1);
1280 tsc.d = dx2d(sc->dx);
1281 tsc.m2 = sm2m(sc->sm2);
1282 if (nla_put(skb, attr, sizeof(tsc), &tsc))
1283 goto nla_put_failure;
1292 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1294 if ((cl->cl_flags & HFSC_RSC) &&
1295 (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1296 goto nla_put_failure;
1298 if ((cl->cl_flags & HFSC_FSC) &&
1299 (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1300 goto nla_put_failure;
1302 if ((cl->cl_flags & HFSC_USC) &&
1303 (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1304 goto nla_put_failure;
1313 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1316 struct hfsc_class *cl = (struct hfsc_class *)arg;
1317 struct nlattr *nest;
1319 tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid :
1321 tcm->tcm_handle = cl->cl_common.classid;
1323 tcm->tcm_info = cl->qdisc->handle;
1325 nest = nla_nest_start(skb, TCA_OPTIONS);
1327 goto nla_put_failure;
1328 if (hfsc_dump_curves(skb, cl) < 0)
1329 goto nla_put_failure;
1330 return nla_nest_end(skb, nest);
1333 nla_nest_cancel(skb, nest);
1338 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1339 struct gnet_dump *d)
1341 struct hfsc_class *cl = (struct hfsc_class *)arg;
1342 struct tc_hfsc_stats xstats;
1344 cl->qstats.backlog = cl->qdisc->qstats.backlog;
1345 xstats.level = cl->level;
1346 xstats.period = cl->cl_vtperiod;
1347 xstats.work = cl->cl_total;
1348 xstats.rtwork = cl->cl_cumul;
1350 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch), d, NULL, &cl->bstats) < 0 ||
1351 gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
1352 gnet_stats_copy_queue(d, NULL, &cl->qstats, cl->qdisc->q.qlen) < 0)
1355 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1361 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1363 struct hfsc_sched *q = qdisc_priv(sch);
1364 struct hfsc_class *cl;
1370 for (i = 0; i < q->clhash.hashsize; i++) {
1371 hlist_for_each_entry(cl, &q->clhash.hash[i],
1373 if (arg->count < arg->skip) {
1377 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1387 hfsc_schedule_watchdog(struct Qdisc *sch)
1389 struct hfsc_sched *q = qdisc_priv(sch);
1390 struct hfsc_class *cl;
1393 cl = eltree_get_minel(q);
1395 next_time = cl->cl_e;
1396 if (q->root.cl_cfmin != 0) {
1397 if (next_time == 0 || next_time > q->root.cl_cfmin)
1398 next_time = q->root.cl_cfmin;
1400 WARN_ON(next_time == 0);
1401 qdisc_watchdog_schedule(&q->watchdog, next_time);
1405 hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1407 struct hfsc_sched *q = qdisc_priv(sch);
1408 struct tc_hfsc_qopt *qopt;
1411 if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1413 qopt = nla_data(opt);
1415 q->defcls = qopt->defcls;
1416 err = qdisc_class_hash_init(&q->clhash);
1419 q->eligible = RB_ROOT;
1421 q->root.cl_common.classid = sch->handle;
1424 q->root.qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1426 if (q->root.qdisc == NULL)
1427 q->root.qdisc = &noop_qdisc;
1428 INIT_LIST_HEAD(&q->root.children);
1429 q->root.vt_tree = RB_ROOT;
1430 q->root.cf_tree = RB_ROOT;
1432 qdisc_class_hash_insert(&q->clhash, &q->root.cl_common);
1433 qdisc_class_hash_grow(sch, &q->clhash);
1435 qdisc_watchdog_init(&q->watchdog, sch);
1441 hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt)
1443 struct hfsc_sched *q = qdisc_priv(sch);
1444 struct tc_hfsc_qopt *qopt;
1446 if (opt == NULL || nla_len(opt) < sizeof(*qopt))
1448 qopt = nla_data(opt);
1451 q->defcls = qopt->defcls;
1452 sch_tree_unlock(sch);
1458 hfsc_reset_class(struct hfsc_class *cl)
1468 cl->cl_vtperiod = 0;
1469 cl->cl_parentperiod = 0;
1475 cl->vt_tree = RB_ROOT;
1476 cl->cf_tree = RB_ROOT;
1477 qdisc_reset(cl->qdisc);
1479 if (cl->cl_flags & HFSC_RSC)
1480 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1481 if (cl->cl_flags & HFSC_FSC)
1482 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1483 if (cl->cl_flags & HFSC_USC)
1484 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1488 hfsc_reset_qdisc(struct Qdisc *sch)
1490 struct hfsc_sched *q = qdisc_priv(sch);
1491 struct hfsc_class *cl;
1494 for (i = 0; i < q->clhash.hashsize; i++) {
1495 hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1496 hfsc_reset_class(cl);
1498 q->eligible = RB_ROOT;
1499 qdisc_watchdog_cancel(&q->watchdog);
1500 sch->qstats.backlog = 0;
1505 hfsc_destroy_qdisc(struct Qdisc *sch)
1507 struct hfsc_sched *q = qdisc_priv(sch);
1508 struct hlist_node *next;
1509 struct hfsc_class *cl;
1512 for (i = 0; i < q->clhash.hashsize; i++) {
1513 hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1514 tcf_destroy_chain(&cl->filter_list);
1516 for (i = 0; i < q->clhash.hashsize; i++) {
1517 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1519 hfsc_destroy_class(sch, cl);
1521 qdisc_class_hash_destroy(&q->clhash);
1522 qdisc_watchdog_cancel(&q->watchdog);
1526 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1528 struct hfsc_sched *q = qdisc_priv(sch);
1529 unsigned char *b = skb_tail_pointer(skb);
1530 struct tc_hfsc_qopt qopt;
1532 qopt.defcls = q->defcls;
1533 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1534 goto nla_put_failure;
1543 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free)
1545 struct hfsc_class *cl;
1546 int uninitialized_var(err);
1548 cl = hfsc_classify(skb, sch, &err);
1550 if (err & __NET_XMIT_BYPASS)
1551 qdisc_qstats_drop(sch);
1552 __qdisc_drop(skb, to_free);
1556 err = qdisc_enqueue(skb, cl->qdisc, to_free);
1557 if (unlikely(err != NET_XMIT_SUCCESS)) {
1558 if (net_xmit_drop_count(err)) {
1560 qdisc_qstats_drop(sch);
1565 if (cl->qdisc->q.qlen == 1) {
1566 set_active(cl, qdisc_pkt_len(skb));
1568 * If this is the first packet, isolate the head so an eventual
1569 * head drop before the first dequeue operation has no chance
1570 * to invalidate the deadline.
1572 if (cl->cl_flags & HFSC_RSC)
1573 cl->qdisc->ops->peek(cl->qdisc);
1577 qdisc_qstats_backlog_inc(sch, skb);
1580 return NET_XMIT_SUCCESS;
1583 static struct sk_buff *
1584 hfsc_dequeue(struct Qdisc *sch)
1586 struct hfsc_sched *q = qdisc_priv(sch);
1587 struct hfsc_class *cl;
1588 struct sk_buff *skb;
1590 unsigned int next_len;
1593 if (sch->q.qlen == 0)
1596 cur_time = psched_get_time();
1599 * if there are eligible classes, use real-time criteria.
1600 * find the class with the minimum deadline among
1601 * the eligible classes.
1603 cl = eltree_get_mindl(q, cur_time);
1608 * use link-sharing criteria
1609 * get the class with the minimum vt in the hierarchy
1611 cl = vttree_get_minvt(&q->root, cur_time);
1613 qdisc_qstats_overlimit(sch);
1614 hfsc_schedule_watchdog(sch);
1619 skb = qdisc_dequeue_peeked(cl->qdisc);
1621 qdisc_warn_nonwc("HFSC", cl->qdisc);
1625 bstats_update(&cl->bstats, skb);
1626 update_vf(cl, qdisc_pkt_len(skb), cur_time);
1628 cl->cl_cumul += qdisc_pkt_len(skb);
1630 if (cl->qdisc->q.qlen != 0) {
1631 if (cl->cl_flags & HFSC_RSC) {
1633 next_len = qdisc_peek_len(cl->qdisc);
1635 update_ed(cl, next_len);
1637 update_d(cl, next_len);
1640 /* the class becomes passive */
1644 qdisc_bstats_update(sch, skb);
1645 qdisc_qstats_backlog_dec(sch, skb);
1651 static const struct Qdisc_class_ops hfsc_class_ops = {
1652 .change = hfsc_change_class,
1653 .delete = hfsc_delete_class,
1654 .graft = hfsc_graft_class,
1655 .leaf = hfsc_class_leaf,
1656 .qlen_notify = hfsc_qlen_notify,
1657 .get = hfsc_get_class,
1658 .put = hfsc_put_class,
1659 .bind_tcf = hfsc_bind_tcf,
1660 .unbind_tcf = hfsc_unbind_tcf,
1661 .tcf_chain = hfsc_tcf_chain,
1662 .dump = hfsc_dump_class,
1663 .dump_stats = hfsc_dump_class_stats,
1667 static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1669 .init = hfsc_init_qdisc,
1670 .change = hfsc_change_qdisc,
1671 .reset = hfsc_reset_qdisc,
1672 .destroy = hfsc_destroy_qdisc,
1673 .dump = hfsc_dump_qdisc,
1674 .enqueue = hfsc_enqueue,
1675 .dequeue = hfsc_dequeue,
1676 .peek = qdisc_peek_dequeued,
1677 .cl_ops = &hfsc_class_ops,
1678 .priv_size = sizeof(struct hfsc_sched),
1679 .owner = THIS_MODULE
1685 return register_qdisc(&hfsc_qdisc_ops);
1691 unregister_qdisc(&hfsc_qdisc_ops);
1694 MODULE_LICENSE("GPL");
1695 module_init(hfsc_init);
1696 module_exit(hfsc_cleanup);