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 u32 classid; /* class id */
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 hlist; /* hash list member */
138 struct list_head dlist; /* drop list member */
140 u64 cl_total; /* total work in bytes */
141 u64 cl_cumul; /* cumulative work in bytes done by
142 real-time criteria */
144 u64 cl_d; /* deadline*/
145 u64 cl_e; /* eligible time */
146 u64 cl_vt; /* virtual time */
147 u64 cl_f; /* time when this class will fit for
148 link-sharing, max(myf, cfmin) */
149 u64 cl_myf; /* my fit-time (calculated from this
150 class's own upperlimit curve) */
151 u64 cl_myfadj; /* my fit-time adjustment (to cancel
152 history dependence) */
153 u64 cl_cfmin; /* earliest children's fit-time (used
154 with cl_myf to obtain cl_f) */
155 u64 cl_cvtmin; /* minimal virtual time among the
156 children fit for link-sharing
157 (monotonic within a period) */
158 u64 cl_vtadj; /* intra-period cumulative vt
160 u64 cl_vtoff; /* inter-period cumulative vt offset */
161 u64 cl_cvtmax; /* max child's vt in the last period */
162 u64 cl_cvtoff; /* cumulative cvtmax of all periods */
163 u64 cl_pcvtoff; /* parent's cvtoff at initialization
166 struct internal_sc cl_rsc; /* internal real-time service curve */
167 struct internal_sc cl_fsc; /* internal fair service curve */
168 struct internal_sc cl_usc; /* internal upperlimit service curve */
169 struct runtime_sc cl_deadline; /* deadline curve */
170 struct runtime_sc cl_eligible; /* eligible curve */
171 struct runtime_sc cl_virtual; /* virtual curve */
172 struct runtime_sc cl_ulimit; /* upperlimit curve */
174 unsigned long cl_flags; /* which curves are valid */
175 unsigned long cl_vtperiod; /* vt period sequence number */
176 unsigned long cl_parentperiod;/* parent's vt period sequence number*/
177 unsigned long cl_nactive; /* number of active children */
180 #define HFSC_HSIZE 16
184 u16 defcls; /* default class id */
185 struct hfsc_class root; /* root class */
186 struct list_head clhash[HFSC_HSIZE]; /* class hash */
187 struct rb_root eligible; /* eligible tree */
188 struct list_head droplist; /* active leaf class list (for
190 struct sk_buff_head requeue; /* requeued packet */
191 struct qdisc_watchdog watchdog; /* watchdog timer */
194 #define HT_INFINITY 0xffffffffffffffffULL /* infinite time value */
198 * eligible tree holds backlogged classes being sorted by their eligible times.
199 * there is one eligible tree per hfsc instance.
203 eltree_insert(struct hfsc_class *cl)
205 struct rb_node **p = &cl->sched->eligible.rb_node;
206 struct rb_node *parent = NULL;
207 struct hfsc_class *cl1;
211 cl1 = rb_entry(parent, struct hfsc_class, el_node);
212 if (cl->cl_e >= cl1->cl_e)
213 p = &parent->rb_right;
215 p = &parent->rb_left;
217 rb_link_node(&cl->el_node, parent, p);
218 rb_insert_color(&cl->el_node, &cl->sched->eligible);
222 eltree_remove(struct hfsc_class *cl)
224 rb_erase(&cl->el_node, &cl->sched->eligible);
228 eltree_update(struct hfsc_class *cl)
234 /* find the class with the minimum deadline among the eligible classes */
235 static inline struct hfsc_class *
236 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
238 struct hfsc_class *p, *cl = NULL;
241 for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
242 p = rb_entry(n, struct hfsc_class, el_node);
243 if (p->cl_e > cur_time)
245 if (cl == NULL || p->cl_d < cl->cl_d)
251 /* find the class with minimum eligible time among the eligible classes */
252 static inline struct hfsc_class *
253 eltree_get_minel(struct hfsc_sched *q)
257 n = rb_first(&q->eligible);
260 return rb_entry(n, struct hfsc_class, el_node);
264 * vttree holds holds backlogged child classes being sorted by their virtual
265 * time. each intermediate class has one vttree.
268 vttree_insert(struct hfsc_class *cl)
270 struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
271 struct rb_node *parent = NULL;
272 struct hfsc_class *cl1;
276 cl1 = rb_entry(parent, struct hfsc_class, vt_node);
277 if (cl->cl_vt >= cl1->cl_vt)
278 p = &parent->rb_right;
280 p = &parent->rb_left;
282 rb_link_node(&cl->vt_node, parent, p);
283 rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
287 vttree_remove(struct hfsc_class *cl)
289 rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
293 vttree_update(struct hfsc_class *cl)
299 static inline struct hfsc_class *
300 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
302 struct hfsc_class *p;
305 for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
306 p = rb_entry(n, struct hfsc_class, vt_node);
307 if (p->cl_f <= cur_time)
314 * get the leaf class with the minimum vt in the hierarchy
316 static struct hfsc_class *
317 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
319 /* if root-class's cfmin is bigger than cur_time nothing to do */
320 if (cl->cl_cfmin > cur_time)
323 while (cl->level > 0) {
324 cl = vttree_firstfit(cl, cur_time);
328 * update parent's cl_cvtmin.
330 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
331 cl->cl_parent->cl_cvtmin = cl->cl_vt;
337 cftree_insert(struct hfsc_class *cl)
339 struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
340 struct rb_node *parent = NULL;
341 struct hfsc_class *cl1;
345 cl1 = rb_entry(parent, struct hfsc_class, cf_node);
346 if (cl->cl_f >= cl1->cl_f)
347 p = &parent->rb_right;
349 p = &parent->rb_left;
351 rb_link_node(&cl->cf_node, parent, p);
352 rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
356 cftree_remove(struct hfsc_class *cl)
358 rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
362 cftree_update(struct hfsc_class *cl)
369 * service curve support functions
371 * external service curve parameters
374 * internal service curve parameters
375 * sm: (bytes/psched_us) << SM_SHIFT
376 * ism: (psched_us/byte) << ISM_SHIFT
379 * The clock source resolution with ktime is 1.024us.
381 * sm and ism are scaled in order to keep effective digits.
382 * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
383 * digits in decimal using the following table.
385 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
386 * ------------+-------------------------------------------------------
387 * bytes/1.024us 12.8e-3 128e-3 1280e-3 12800e-3 128000e-3
389 * 1.024us/byte 78.125 7.8125 0.78125 0.078125 0.0078125
394 #define SM_MASK ((1ULL << SM_SHIFT) - 1)
395 #define ISM_MASK ((1ULL << ISM_SHIFT) - 1)
398 seg_x2y(u64 x, u64 sm)
404 * y = x * sm >> SM_SHIFT
405 * but divide it for the upper and lower bits to avoid overflow
407 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
412 seg_y2x(u64 y, u64 ism)
418 else if (ism == HT_INFINITY)
421 x = (y >> ISM_SHIFT) * ism
422 + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
427 /* Convert m (bps) into sm (bytes/psched us) */
433 sm = ((u64)m << SM_SHIFT);
434 sm += PSCHED_TICKS_PER_SEC - 1;
435 do_div(sm, PSCHED_TICKS_PER_SEC);
439 /* convert m (bps) into ism (psched us/byte) */
448 ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
455 /* convert d (us) into dx (psched us) */
461 dx = ((u64)d * PSCHED_TICKS_PER_SEC);
462 dx += USEC_PER_SEC - 1;
463 do_div(dx, USEC_PER_SEC);
467 /* convert sm (bytes/psched us) into m (bps) */
473 m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
477 /* convert dx (psched us) into d (us) */
483 d = dx * USEC_PER_SEC;
484 do_div(d, PSCHED_TICKS_PER_SEC);
489 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
491 isc->sm1 = m2sm(sc->m1);
492 isc->ism1 = m2ism(sc->m1);
493 isc->dx = d2dx(sc->d);
494 isc->dy = seg_x2y(isc->dx, isc->sm1);
495 isc->sm2 = m2sm(sc->m2);
496 isc->ism2 = m2ism(sc->m2);
500 * initialize the runtime service curve with the given internal
501 * service curve starting at (x, y).
504 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
508 rtsc->sm1 = isc->sm1;
509 rtsc->ism1 = isc->ism1;
512 rtsc->sm2 = isc->sm2;
513 rtsc->ism2 = isc->ism2;
517 * calculate the y-projection of the runtime service curve by the
518 * given x-projection value
521 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
527 else if (y <= rtsc->y + rtsc->dy) {
528 /* x belongs to the 1st segment */
530 x = rtsc->x + rtsc->dx;
532 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
534 /* x belongs to the 2nd segment */
535 x = rtsc->x + rtsc->dx
536 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
542 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
548 else if (x <= rtsc->x + rtsc->dx)
549 /* y belongs to the 1st segment */
550 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
552 /* y belongs to the 2nd segment */
553 y = rtsc->y + rtsc->dy
554 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
559 * update the runtime service curve by taking the minimum of the current
560 * runtime service curve and the service curve starting at (x, y).
563 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
568 if (isc->sm1 <= isc->sm2) {
569 /* service curve is convex */
570 y1 = rtsc_x2y(rtsc, x);
572 /* the current rtsc is smaller */
580 * service curve is concave
581 * compute the two y values of the current rtsc
585 y1 = rtsc_x2y(rtsc, x);
587 /* rtsc is below isc, no change to rtsc */
591 y2 = rtsc_x2y(rtsc, x + isc->dx);
592 if (y2 >= y + isc->dy) {
593 /* rtsc is above isc, replace rtsc by isc */
602 * the two curves intersect
603 * compute the offsets (dx, dy) using the reverse
604 * function of seg_x2y()
605 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
607 dx = (y1 - y) << SM_SHIFT;
608 dsm = isc->sm1 - isc->sm2;
611 * check if (x, y1) belongs to the 1st segment of rtsc.
612 * if so, add the offset.
614 if (rtsc->x + rtsc->dx > x)
615 dx += rtsc->x + rtsc->dx - x;
616 dy = seg_x2y(dx, isc->sm1);
626 init_ed(struct hfsc_class *cl, unsigned int next_len)
628 u64 cur_time = psched_get_time();
630 /* update the deadline curve */
631 rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
634 * update the eligible curve.
635 * for concave, it is equal to the deadline curve.
636 * for convex, it is a linear curve with slope m2.
638 cl->cl_eligible = cl->cl_deadline;
639 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
640 cl->cl_eligible.dx = 0;
641 cl->cl_eligible.dy = 0;
644 /* compute e and d */
645 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
646 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
652 update_ed(struct hfsc_class *cl, unsigned int next_len)
654 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
655 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
661 update_d(struct hfsc_class *cl, unsigned int next_len)
663 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
667 update_cfmin(struct hfsc_class *cl)
669 struct rb_node *n = rb_first(&cl->cf_tree);
670 struct hfsc_class *p;
676 p = rb_entry(n, struct hfsc_class, cf_node);
677 cl->cl_cfmin = p->cl_f;
681 init_vf(struct hfsc_class *cl, unsigned int len)
683 struct hfsc_class *max_cl;
690 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
691 if (go_active && cl->cl_nactive++ == 0)
697 n = rb_last(&cl->cl_parent->vt_tree);
699 max_cl = rb_entry(n, struct hfsc_class,vt_node);
701 * set vt to the average of the min and max
702 * classes. if the parent's period didn't
703 * change, don't decrease vt of the class.
706 if (cl->cl_parent->cl_cvtmin != 0)
707 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
709 if (cl->cl_parent->cl_vtperiod !=
710 cl->cl_parentperiod || vt > cl->cl_vt)
714 * first child for a new parent backlog period.
715 * add parent's cvtmax to cvtoff to make a new
716 * vt (vtoff + vt) larger than the vt in the
717 * last period for all children.
719 vt = cl->cl_parent->cl_cvtmax;
720 cl->cl_parent->cl_cvtoff += vt;
721 cl->cl_parent->cl_cvtmax = 0;
722 cl->cl_parent->cl_cvtmin = 0;
726 cl->cl_vtoff = cl->cl_parent->cl_cvtoff -
729 /* update the virtual curve */
730 vt = cl->cl_vt + cl->cl_vtoff;
731 rtsc_min(&cl->cl_virtual, &cl->cl_fsc, vt,
733 if (cl->cl_virtual.x == vt) {
734 cl->cl_virtual.x -= cl->cl_vtoff;
739 cl->cl_vtperiod++; /* increment vt period */
740 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
741 if (cl->cl_parent->cl_nactive == 0)
742 cl->cl_parentperiod++;
748 if (cl->cl_flags & HFSC_USC) {
749 /* class has upper limit curve */
751 cur_time = psched_get_time();
753 /* update the ulimit curve */
754 rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
757 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
763 f = max(cl->cl_myf, cl->cl_cfmin);
767 update_cfmin(cl->cl_parent);
773 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
775 u64 f; /* , myf_bound, delta; */
778 if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
781 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
784 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
787 if (go_passive && --cl->cl_nactive == 0)
793 /* no more active child, going passive */
795 /* update cvtmax of the parent class */
796 if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
797 cl->cl_parent->cl_cvtmax = cl->cl_vt;
799 /* remove this class from the vt tree */
803 update_cfmin(cl->cl_parent);
811 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
812 - cl->cl_vtoff + cl->cl_vtadj;
815 * if vt of the class is smaller than cvtmin,
816 * the class was skipped in the past due to non-fit.
817 * if so, we need to adjust vtadj.
819 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
820 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
821 cl->cl_vt = cl->cl_parent->cl_cvtmin;
824 /* update the vt tree */
827 if (cl->cl_flags & HFSC_USC) {
828 cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
832 * This code causes classes to stay way under their
833 * limit when multiple classes are used at gigabit
834 * speed. needs investigation. -kaber
837 * if myf lags behind by more than one clock tick
838 * from the current time, adjust myfadj to prevent
839 * a rate-limited class from going greedy.
840 * in a steady state under rate-limiting, myf
841 * fluctuates within one clock tick.
843 myf_bound = cur_time - PSCHED_JIFFIE2US(1);
844 if (cl->cl_myf < myf_bound) {
845 delta = cur_time - cl->cl_myf;
846 cl->cl_myfadj += delta;
852 f = max(cl->cl_myf, cl->cl_cfmin);
856 update_cfmin(cl->cl_parent);
862 set_active(struct hfsc_class *cl, unsigned int len)
864 if (cl->cl_flags & HFSC_RSC)
866 if (cl->cl_flags & HFSC_FSC)
869 list_add_tail(&cl->dlist, &cl->sched->droplist);
873 set_passive(struct hfsc_class *cl)
875 if (cl->cl_flags & HFSC_RSC)
878 list_del(&cl->dlist);
881 * vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
882 * needs to be called explicitly to remove a class from vttree.
887 * hack to get length of first packet in queue.
890 qdisc_peek_len(struct Qdisc *sch)
895 skb = sch->dequeue(sch);
898 printk("qdisc_peek_len: non work-conserving qdisc ?\n");
902 if (unlikely(sch->ops->requeue(skb, sch) != NET_XMIT_SUCCESS)) {
904 printk("qdisc_peek_len: failed to requeue\n");
905 qdisc_tree_decrease_qlen(sch, 1);
912 hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl)
914 unsigned int len = cl->qdisc->q.qlen;
916 qdisc_reset(cl->qdisc);
917 qdisc_tree_decrease_qlen(cl->qdisc, len);
921 hfsc_adjust_levels(struct hfsc_class *cl)
923 struct hfsc_class *p;
928 list_for_each_entry(p, &cl->children, siblings) {
929 if (p->level >= level)
930 level = p->level + 1;
933 } while ((cl = cl->cl_parent) != NULL);
936 static inline unsigned int
942 return h & (HFSC_HSIZE - 1);
945 static inline struct hfsc_class *
946 hfsc_find_class(u32 classid, struct Qdisc *sch)
948 struct hfsc_sched *q = qdisc_priv(sch);
949 struct hfsc_class *cl;
951 list_for_each_entry(cl, &q->clhash[hfsc_hash(classid)], hlist) {
952 if (cl->classid == classid)
959 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
962 sc2isc(rsc, &cl->cl_rsc);
963 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
964 cl->cl_eligible = cl->cl_deadline;
965 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
966 cl->cl_eligible.dx = 0;
967 cl->cl_eligible.dy = 0;
969 cl->cl_flags |= HFSC_RSC;
973 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
975 sc2isc(fsc, &cl->cl_fsc);
976 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
977 cl->cl_flags |= HFSC_FSC;
981 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
984 sc2isc(usc, &cl->cl_usc);
985 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
986 cl->cl_flags |= HFSC_USC;
990 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
991 struct rtattr **tca, unsigned long *arg)
993 struct hfsc_sched *q = qdisc_priv(sch);
994 struct hfsc_class *cl = (struct hfsc_class *)*arg;
995 struct hfsc_class *parent = NULL;
996 struct rtattr *opt = tca[TCA_OPTIONS-1];
997 struct rtattr *tb[TCA_HFSC_MAX];
998 struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
1001 if (opt == NULL || rtattr_parse_nested(tb, TCA_HFSC_MAX, opt))
1004 if (tb[TCA_HFSC_RSC-1]) {
1005 if (RTA_PAYLOAD(tb[TCA_HFSC_RSC-1]) < sizeof(*rsc))
1007 rsc = RTA_DATA(tb[TCA_HFSC_RSC-1]);
1008 if (rsc->m1 == 0 && rsc->m2 == 0)
1012 if (tb[TCA_HFSC_FSC-1]) {
1013 if (RTA_PAYLOAD(tb[TCA_HFSC_FSC-1]) < sizeof(*fsc))
1015 fsc = RTA_DATA(tb[TCA_HFSC_FSC-1]);
1016 if (fsc->m1 == 0 && fsc->m2 == 0)
1020 if (tb[TCA_HFSC_USC-1]) {
1021 if (RTA_PAYLOAD(tb[TCA_HFSC_USC-1]) < sizeof(*usc))
1023 usc = RTA_DATA(tb[TCA_HFSC_USC-1]);
1024 if (usc->m1 == 0 && usc->m2 == 0)
1030 if (cl->cl_parent && cl->cl_parent->classid != parentid)
1032 if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
1035 cur_time = psched_get_time();
1039 hfsc_change_rsc(cl, rsc, cur_time);
1041 hfsc_change_fsc(cl, fsc);
1043 hfsc_change_usc(cl, usc, cur_time);
1045 if (cl->qdisc->q.qlen != 0) {
1046 if (cl->cl_flags & HFSC_RSC)
1047 update_ed(cl, qdisc_peek_len(cl->qdisc));
1048 if (cl->cl_flags & HFSC_FSC)
1049 update_vf(cl, 0, cur_time);
1051 sch_tree_unlock(sch);
1053 if (tca[TCA_RATE-1])
1054 gen_replace_estimator(&cl->bstats, &cl->rate_est,
1055 &sch->dev->queue_lock,
1060 if (parentid == TC_H_ROOT)
1065 parent = hfsc_find_class(parentid, sch);
1070 if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1072 if (hfsc_find_class(classid, sch))
1075 if (rsc == NULL && fsc == NULL)
1078 cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1083 hfsc_change_rsc(cl, rsc, 0);
1085 hfsc_change_fsc(cl, fsc);
1087 hfsc_change_usc(cl, usc, 0);
1090 cl->classid = classid;
1092 cl->cl_parent = parent;
1093 cl->qdisc = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops, classid);
1094 if (cl->qdisc == NULL)
1095 cl->qdisc = &noop_qdisc;
1096 INIT_LIST_HEAD(&cl->children);
1097 cl->vt_tree = RB_ROOT;
1098 cl->cf_tree = RB_ROOT;
1101 list_add_tail(&cl->hlist, &q->clhash[hfsc_hash(classid)]);
1102 list_add_tail(&cl->siblings, &parent->children);
1103 if (parent->level == 0)
1104 hfsc_purge_queue(sch, parent);
1105 hfsc_adjust_levels(parent);
1106 cl->cl_pcvtoff = parent->cl_cvtoff;
1107 sch_tree_unlock(sch);
1109 if (tca[TCA_RATE-1])
1110 gen_new_estimator(&cl->bstats, &cl->rate_est,
1111 &sch->dev->queue_lock, tca[TCA_RATE-1]);
1112 *arg = (unsigned long)cl;
1117 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1119 struct hfsc_sched *q = qdisc_priv(sch);
1121 tcf_destroy_chain(cl->filter_list);
1122 qdisc_destroy(cl->qdisc);
1123 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1129 hfsc_delete_class(struct Qdisc *sch, unsigned long arg)
1131 struct hfsc_sched *q = qdisc_priv(sch);
1132 struct hfsc_class *cl = (struct hfsc_class *)arg;
1134 if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1139 list_del(&cl->siblings);
1140 hfsc_adjust_levels(cl->cl_parent);
1142 hfsc_purge_queue(sch, cl);
1143 list_del(&cl->hlist);
1145 if (--cl->refcnt == 0)
1146 hfsc_destroy_class(sch, cl);
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_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;
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(sch->dev, &pfifo_qdisc_ops,
1216 hfsc_purge_queue(sch, cl);
1217 *old = xchg(&cl->qdisc, new);
1218 sch_tree_unlock(sch);
1222 static struct Qdisc *
1223 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1225 struct hfsc_class *cl = (struct hfsc_class *)arg;
1227 if (cl != NULL && cl->level == 0)
1234 hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1236 struct hfsc_class *cl = (struct hfsc_class *)arg;
1238 if (cl->qdisc->q.qlen == 0) {
1239 update_vf(cl, 0, 0);
1244 static unsigned long
1245 hfsc_get_class(struct Qdisc *sch, u32 classid)
1247 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1252 return (unsigned long)cl;
1256 hfsc_put_class(struct Qdisc *sch, unsigned long arg)
1258 struct hfsc_class *cl = (struct hfsc_class *)arg;
1260 if (--cl->refcnt == 0)
1261 hfsc_destroy_class(sch, cl);
1264 static unsigned long
1265 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1267 struct hfsc_class *p = (struct hfsc_class *)parent;
1268 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1271 if (p != NULL && p->level <= cl->level)
1276 return (unsigned long)cl;
1280 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1282 struct hfsc_class *cl = (struct hfsc_class *)arg;
1287 static struct tcf_proto **
1288 hfsc_tcf_chain(struct Qdisc *sch, unsigned long arg)
1290 struct hfsc_sched *q = qdisc_priv(sch);
1291 struct hfsc_class *cl = (struct hfsc_class *)arg;
1296 return &cl->filter_list;
1300 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1302 struct tc_service_curve tsc;
1304 tsc.m1 = sm2m(sc->sm1);
1305 tsc.d = dx2d(sc->dx);
1306 tsc.m2 = sm2m(sc->sm2);
1307 RTA_PUT(skb, attr, sizeof(tsc), &tsc);
1316 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1318 if ((cl->cl_flags & HFSC_RSC) &&
1319 (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1320 goto rtattr_failure;
1322 if ((cl->cl_flags & HFSC_FSC) &&
1323 (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1324 goto rtattr_failure;
1326 if ((cl->cl_flags & HFSC_USC) &&
1327 (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1328 goto rtattr_failure;
1337 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1340 struct hfsc_class *cl = (struct hfsc_class *)arg;
1341 unsigned char *b = skb_tail_pointer(skb);
1342 struct rtattr *rta = (struct rtattr *)b;
1344 tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->classid : TC_H_ROOT;
1345 tcm->tcm_handle = cl->classid;
1347 tcm->tcm_info = cl->qdisc->handle;
1349 RTA_PUT(skb, TCA_OPTIONS, 0, NULL);
1350 if (hfsc_dump_curves(skb, cl) < 0)
1351 goto rtattr_failure;
1352 rta->rta_len = skb_tail_pointer(skb) - b;
1361 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1362 struct gnet_dump *d)
1364 struct hfsc_class *cl = (struct hfsc_class *)arg;
1365 struct tc_hfsc_stats xstats;
1367 cl->qstats.qlen = cl->qdisc->q.qlen;
1368 xstats.level = cl->level;
1369 xstats.period = cl->cl_vtperiod;
1370 xstats.work = cl->cl_total;
1371 xstats.rtwork = cl->cl_cumul;
1373 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1374 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1375 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1378 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1384 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1386 struct hfsc_sched *q = qdisc_priv(sch);
1387 struct hfsc_class *cl;
1393 for (i = 0; i < HFSC_HSIZE; i++) {
1394 list_for_each_entry(cl, &q->clhash[i], hlist) {
1395 if (arg->count < arg->skip) {
1399 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1409 hfsc_schedule_watchdog(struct Qdisc *sch)
1411 struct hfsc_sched *q = qdisc_priv(sch);
1412 struct hfsc_class *cl;
1415 if ((cl = eltree_get_minel(q)) != NULL)
1416 next_time = cl->cl_e;
1417 if (q->root.cl_cfmin != 0) {
1418 if (next_time == 0 || next_time > q->root.cl_cfmin)
1419 next_time = q->root.cl_cfmin;
1421 WARN_ON(next_time == 0);
1422 qdisc_watchdog_schedule(&q->watchdog, next_time);
1426 hfsc_init_qdisc(struct Qdisc *sch, struct rtattr *opt)
1428 struct hfsc_sched *q = qdisc_priv(sch);
1429 struct tc_hfsc_qopt *qopt;
1432 if (opt == NULL || RTA_PAYLOAD(opt) < sizeof(*qopt))
1434 qopt = RTA_DATA(opt);
1436 q->defcls = qopt->defcls;
1437 for (i = 0; i < HFSC_HSIZE; i++)
1438 INIT_LIST_HEAD(&q->clhash[i]);
1439 q->eligible = RB_ROOT;
1440 INIT_LIST_HEAD(&q->droplist);
1441 skb_queue_head_init(&q->requeue);
1444 q->root.classid = sch->handle;
1446 q->root.qdisc = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops,
1448 if (q->root.qdisc == NULL)
1449 q->root.qdisc = &noop_qdisc;
1450 INIT_LIST_HEAD(&q->root.children);
1451 q->root.vt_tree = RB_ROOT;
1452 q->root.cf_tree = RB_ROOT;
1454 list_add(&q->root.hlist, &q->clhash[hfsc_hash(q->root.classid)]);
1456 qdisc_watchdog_init(&q->watchdog, sch);
1462 hfsc_change_qdisc(struct Qdisc *sch, struct rtattr *opt)
1464 struct hfsc_sched *q = qdisc_priv(sch);
1465 struct tc_hfsc_qopt *qopt;
1467 if (opt == NULL || RTA_PAYLOAD(opt) < sizeof(*qopt))
1469 qopt = RTA_DATA(opt);
1472 q->defcls = qopt->defcls;
1473 sch_tree_unlock(sch);
1479 hfsc_reset_class(struct hfsc_class *cl)
1492 cl->cl_vtperiod = 0;
1493 cl->cl_parentperiod = 0;
1500 cl->vt_tree = RB_ROOT;
1501 cl->cf_tree = RB_ROOT;
1502 qdisc_reset(cl->qdisc);
1504 if (cl->cl_flags & HFSC_RSC)
1505 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1506 if (cl->cl_flags & HFSC_FSC)
1507 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1508 if (cl->cl_flags & HFSC_USC)
1509 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1513 hfsc_reset_qdisc(struct Qdisc *sch)
1515 struct hfsc_sched *q = qdisc_priv(sch);
1516 struct hfsc_class *cl;
1519 for (i = 0; i < HFSC_HSIZE; i++) {
1520 list_for_each_entry(cl, &q->clhash[i], hlist)
1521 hfsc_reset_class(cl);
1523 __skb_queue_purge(&q->requeue);
1524 q->eligible = RB_ROOT;
1525 INIT_LIST_HEAD(&q->droplist);
1526 qdisc_watchdog_cancel(&q->watchdog);
1531 hfsc_destroy_qdisc(struct Qdisc *sch)
1533 struct hfsc_sched *q = qdisc_priv(sch);
1534 struct hfsc_class *cl, *next;
1537 for (i = 0; i < HFSC_HSIZE; i++) {
1538 list_for_each_entry_safe(cl, next, &q->clhash[i], hlist)
1539 hfsc_destroy_class(sch, cl);
1541 __skb_queue_purge(&q->requeue);
1542 qdisc_watchdog_cancel(&q->watchdog);
1546 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1548 struct hfsc_sched *q = qdisc_priv(sch);
1549 unsigned char *b = skb_tail_pointer(skb);
1550 struct tc_hfsc_qopt qopt;
1552 qopt.defcls = q->defcls;
1553 RTA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
1562 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1564 struct hfsc_class *cl;
1568 cl = hfsc_classify(skb, sch, &err);
1570 if (err == NET_XMIT_BYPASS)
1571 sch->qstats.drops++;
1577 err = cl->qdisc->enqueue(skb, cl->qdisc);
1578 if (unlikely(err != NET_XMIT_SUCCESS)) {
1580 sch->qstats.drops++;
1584 if (cl->qdisc->q.qlen == 1)
1585 set_active(cl, len);
1587 cl->bstats.packets++;
1588 cl->bstats.bytes += len;
1589 sch->bstats.packets++;
1590 sch->bstats.bytes += len;
1593 return NET_XMIT_SUCCESS;
1596 static struct sk_buff *
1597 hfsc_dequeue(struct Qdisc *sch)
1599 struct hfsc_sched *q = qdisc_priv(sch);
1600 struct hfsc_class *cl;
1601 struct sk_buff *skb;
1603 unsigned int next_len;
1606 if (sch->q.qlen == 0)
1608 if ((skb = __skb_dequeue(&q->requeue)))
1611 cur_time = psched_get_time();
1614 * if there are eligible classes, use real-time criteria.
1615 * find the class with the minimum deadline among
1616 * the eligible classes.
1618 if ((cl = eltree_get_mindl(q, cur_time)) != NULL) {
1622 * use link-sharing criteria
1623 * get the class with the minimum vt in the hierarchy
1625 cl = vttree_get_minvt(&q->root, cur_time);
1627 sch->qstats.overlimits++;
1628 hfsc_schedule_watchdog(sch);
1633 skb = cl->qdisc->dequeue(cl->qdisc);
1635 if (net_ratelimit())
1636 printk("HFSC: Non-work-conserving qdisc ?\n");
1640 update_vf(cl, skb->len, cur_time);
1642 cl->cl_cumul += skb->len;
1644 if (cl->qdisc->q.qlen != 0) {
1645 if (cl->cl_flags & HFSC_RSC) {
1647 next_len = qdisc_peek_len(cl->qdisc);
1649 update_ed(cl, next_len);
1651 update_d(cl, next_len);
1654 /* the class becomes passive */
1659 sch->flags &= ~TCQ_F_THROTTLED;
1666 hfsc_requeue(struct sk_buff *skb, struct Qdisc *sch)
1668 struct hfsc_sched *q = qdisc_priv(sch);
1670 __skb_queue_head(&q->requeue, skb);
1672 sch->qstats.requeues++;
1673 return NET_XMIT_SUCCESS;
1677 hfsc_drop(struct Qdisc *sch)
1679 struct hfsc_sched *q = qdisc_priv(sch);
1680 struct hfsc_class *cl;
1683 list_for_each_entry(cl, &q->droplist, dlist) {
1684 if (cl->qdisc->ops->drop != NULL &&
1685 (len = cl->qdisc->ops->drop(cl->qdisc)) > 0) {
1686 if (cl->qdisc->q.qlen == 0) {
1687 update_vf(cl, 0, 0);
1690 list_move_tail(&cl->dlist, &q->droplist);
1693 sch->qstats.drops++;
1701 static const struct Qdisc_class_ops hfsc_class_ops = {
1702 .change = hfsc_change_class,
1703 .delete = hfsc_delete_class,
1704 .graft = hfsc_graft_class,
1705 .leaf = hfsc_class_leaf,
1706 .qlen_notify = hfsc_qlen_notify,
1707 .get = hfsc_get_class,
1708 .put = hfsc_put_class,
1709 .bind_tcf = hfsc_bind_tcf,
1710 .unbind_tcf = hfsc_unbind_tcf,
1711 .tcf_chain = hfsc_tcf_chain,
1712 .dump = hfsc_dump_class,
1713 .dump_stats = hfsc_dump_class_stats,
1717 static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1719 .init = hfsc_init_qdisc,
1720 .change = hfsc_change_qdisc,
1721 .reset = hfsc_reset_qdisc,
1722 .destroy = hfsc_destroy_qdisc,
1723 .dump = hfsc_dump_qdisc,
1724 .enqueue = hfsc_enqueue,
1725 .dequeue = hfsc_dequeue,
1726 .requeue = hfsc_requeue,
1728 .cl_ops = &hfsc_class_ops,
1729 .priv_size = sizeof(struct hfsc_sched),
1730 .owner = THIS_MODULE
1736 return register_qdisc(&hfsc_qdisc_ops);
1742 unregister_qdisc(&hfsc_qdisc_ops);
1745 MODULE_LICENSE("GPL");
1746 module_init(hfsc_init);
1747 module_exit(hfsc_cleanup);