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/jiffies.h>
57 #include <linux/compiler.h>
58 #include <linux/spinlock.h>
59 #include <linux/skbuff.h>
60 #include <linux/string.h>
61 #include <linux/slab.h>
62 #include <linux/timer.h>
63 #include <linux/list.h>
64 #include <linux/rbtree.h>
65 #include <linux/init.h>
66 #include <linux/netdevice.h>
67 #include <linux/rtnetlink.h>
68 #include <linux/pkt_sched.h>
69 #include <net/pkt_sched.h>
70 #include <net/pkt_cls.h>
71 #include <asm/system.h>
72 #include <asm/div64.h>
75 * kernel internal service curve representation:
76 * coordinates are given by 64 bit unsigned integers.
77 * x-axis: unit is clock count.
78 * y-axis: unit is byte.
80 * The service curve parameters are converted to the internal
81 * representation. The slope values are scaled to avoid overflow.
82 * the inverse slope values as well as the y-projection of the 1st
83 * segment are kept in order to to avoid 64-bit divide operations
84 * that are expensive on 32-bit architectures.
89 u64 sm1; /* scaled slope of the 1st segment */
90 u64 ism1; /* scaled inverse-slope of the 1st segment */
91 u64 dx; /* the x-projection of the 1st segment */
92 u64 dy; /* the y-projection of the 1st segment */
93 u64 sm2; /* scaled slope of the 2nd segment */
94 u64 ism2; /* scaled inverse-slope of the 2nd segment */
97 /* runtime service curve */
100 u64 x; /* current starting position on x-axis */
101 u64 y; /* current starting position on y-axis */
102 u64 sm1; /* scaled slope of the 1st segment */
103 u64 ism1; /* scaled inverse-slope of the 1st segment */
104 u64 dx; /* the x-projection of the 1st segment */
105 u64 dy; /* the y-projection of the 1st segment */
106 u64 sm2; /* scaled slope of the 2nd segment */
107 u64 ism2; /* scaled inverse-slope of the 2nd segment */
110 enum hfsc_class_flags
119 u32 classid; /* class id */
120 unsigned int refcnt; /* usage count */
122 struct gnet_stats_basic bstats;
123 struct gnet_stats_queue qstats;
124 struct gnet_stats_rate_est rate_est;
125 spinlock_t *stats_lock;
126 unsigned int level; /* class level in hierarchy */
127 struct tcf_proto *filter_list; /* filter list */
128 unsigned int filter_cnt; /* filter count */
130 struct hfsc_sched *sched; /* scheduler data */
131 struct hfsc_class *cl_parent; /* parent class */
132 struct list_head siblings; /* sibling classes */
133 struct list_head children; /* child classes */
134 struct Qdisc *qdisc; /* leaf qdisc */
136 struct rb_node el_node; /* qdisc's eligible tree member */
137 struct rb_root vt_tree; /* active children sorted by cl_vt */
138 struct rb_node vt_node; /* parent's vt_tree member */
139 struct rb_root cf_tree; /* active children sorted by cl_f */
140 struct rb_node cf_node; /* parent's cf_heap member */
141 struct list_head hlist; /* hash list member */
142 struct list_head dlist; /* drop list member */
144 u64 cl_total; /* total work in bytes */
145 u64 cl_cumul; /* cumulative work in bytes done by
146 real-time criteria */
148 u64 cl_d; /* deadline*/
149 u64 cl_e; /* eligible time */
150 u64 cl_vt; /* virtual time */
151 u64 cl_f; /* time when this class will fit for
152 link-sharing, max(myf, cfmin) */
153 u64 cl_myf; /* my fit-time (calculated from this
154 class's own upperlimit curve) */
155 u64 cl_myfadj; /* my fit-time adjustment (to cancel
156 history dependence) */
157 u64 cl_cfmin; /* earliest children's fit-time (used
158 with cl_myf to obtain cl_f) */
159 u64 cl_cvtmin; /* minimal virtual time among the
160 children fit for link-sharing
161 (monotonic within a period) */
162 u64 cl_vtadj; /* intra-period cumulative vt
164 u64 cl_vtoff; /* inter-period cumulative vt offset */
165 u64 cl_cvtmax; /* max child's vt in the last period */
166 u64 cl_cvtoff; /* cumulative cvtmax of all periods */
167 u64 cl_pcvtoff; /* parent's cvtoff at initalization
170 struct internal_sc cl_rsc; /* internal real-time service curve */
171 struct internal_sc cl_fsc; /* internal fair service curve */
172 struct internal_sc cl_usc; /* internal upperlimit service curve */
173 struct runtime_sc cl_deadline; /* deadline curve */
174 struct runtime_sc cl_eligible; /* eligible curve */
175 struct runtime_sc cl_virtual; /* virtual curve */
176 struct runtime_sc cl_ulimit; /* upperlimit curve */
178 unsigned long cl_flags; /* which curves are valid */
179 unsigned long cl_vtperiod; /* vt period sequence number */
180 unsigned long cl_parentperiod;/* parent's vt period sequence number*/
181 unsigned long cl_nactive; /* number of active children */
184 #define HFSC_HSIZE 16
188 u16 defcls; /* default class id */
189 struct hfsc_class root; /* root class */
190 struct list_head clhash[HFSC_HSIZE]; /* class hash */
191 struct rb_root eligible; /* eligible tree */
192 struct list_head droplist; /* active leaf class list (for
194 struct sk_buff_head requeue; /* requeued packet */
195 struct timer_list wd_timer; /* watchdog timer */
198 #define HT_INFINITY 0xffffffffffffffffULL /* infinite time value */
202 * eligible tree holds backlogged classes being sorted by their eligible times.
203 * there is one eligible tree per hfsc instance.
207 eltree_insert(struct hfsc_class *cl)
209 struct rb_node **p = &cl->sched->eligible.rb_node;
210 struct rb_node *parent = NULL;
211 struct hfsc_class *cl1;
215 cl1 = rb_entry(parent, struct hfsc_class, el_node);
216 if (cl->cl_e >= cl1->cl_e)
217 p = &parent->rb_right;
219 p = &parent->rb_left;
221 rb_link_node(&cl->el_node, parent, p);
222 rb_insert_color(&cl->el_node, &cl->sched->eligible);
226 eltree_remove(struct hfsc_class *cl)
228 rb_erase(&cl->el_node, &cl->sched->eligible);
232 eltree_update(struct hfsc_class *cl)
238 /* find the class with the minimum deadline among the eligible classes */
239 static inline struct hfsc_class *
240 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
242 struct hfsc_class *p, *cl = NULL;
245 for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
246 p = rb_entry(n, struct hfsc_class, el_node);
247 if (p->cl_e > cur_time)
249 if (cl == NULL || p->cl_d < cl->cl_d)
255 /* find the class with minimum eligible time among the eligible classes */
256 static inline struct hfsc_class *
257 eltree_get_minel(struct hfsc_sched *q)
261 n = rb_first(&q->eligible);
264 return rb_entry(n, struct hfsc_class, el_node);
268 * vttree holds holds backlogged child classes being sorted by their virtual
269 * time. each intermediate class has one vttree.
272 vttree_insert(struct hfsc_class *cl)
274 struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
275 struct rb_node *parent = NULL;
276 struct hfsc_class *cl1;
280 cl1 = rb_entry(parent, struct hfsc_class, vt_node);
281 if (cl->cl_vt >= cl1->cl_vt)
282 p = &parent->rb_right;
284 p = &parent->rb_left;
286 rb_link_node(&cl->vt_node, parent, p);
287 rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
291 vttree_remove(struct hfsc_class *cl)
293 rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
297 vttree_update(struct hfsc_class *cl)
303 static inline struct hfsc_class *
304 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
306 struct hfsc_class *p;
309 for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
310 p = rb_entry(n, struct hfsc_class, vt_node);
311 if (p->cl_f <= cur_time)
318 * get the leaf class with the minimum vt in the hierarchy
320 static struct hfsc_class *
321 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
323 /* if root-class's cfmin is bigger than cur_time nothing to do */
324 if (cl->cl_cfmin > cur_time)
327 while (cl->level > 0) {
328 cl = vttree_firstfit(cl, cur_time);
332 * update parent's cl_cvtmin.
334 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
335 cl->cl_parent->cl_cvtmin = cl->cl_vt;
341 cftree_insert(struct hfsc_class *cl)
343 struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
344 struct rb_node *parent = NULL;
345 struct hfsc_class *cl1;
349 cl1 = rb_entry(parent, struct hfsc_class, cf_node);
350 if (cl->cl_f >= cl1->cl_f)
351 p = &parent->rb_right;
353 p = &parent->rb_left;
355 rb_link_node(&cl->cf_node, parent, p);
356 rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
360 cftree_remove(struct hfsc_class *cl)
362 rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
366 cftree_update(struct hfsc_class *cl)
373 * service curve support functions
375 * external service curve parameters
378 * internal service curve parameters
379 * sm: (bytes/psched_us) << SM_SHIFT
380 * ism: (psched_us/byte) << ISM_SHIFT
383 * The clock source resolution with ktime is 1.024us.
385 * sm and ism are scaled in order to keep effective digits.
386 * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
387 * digits in decimal using the following table.
389 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
390 * ------------+-------------------------------------------------------
391 * bytes/1.024us 12.8e-3 128e-3 1280e-3 12800e-3 128000e-3
393 * 1.024us/byte 78.125 7.8125 0.78125 0.078125 0.0078125
398 #define SM_MASK ((1ULL << SM_SHIFT) - 1)
399 #define ISM_MASK ((1ULL << ISM_SHIFT) - 1)
402 seg_x2y(u64 x, u64 sm)
408 * y = x * sm >> SM_SHIFT
409 * but divide it for the upper and lower bits to avoid overflow
411 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
416 seg_y2x(u64 y, u64 ism)
422 else if (ism == HT_INFINITY)
425 x = (y >> ISM_SHIFT) * ism
426 + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
431 /* Convert m (bps) into sm (bytes/psched us) */
437 sm = ((u64)m << SM_SHIFT);
438 sm += PSCHED_JIFFIE2US(HZ) - 1;
439 do_div(sm, PSCHED_JIFFIE2US(HZ));
443 /* convert m (bps) into ism (psched us/byte) */
452 ism = ((u64)PSCHED_JIFFIE2US(HZ) << ISM_SHIFT);
459 /* convert d (us) into dx (psched us) */
465 dx = ((u64)d * PSCHED_JIFFIE2US(HZ));
466 dx += USEC_PER_SEC - 1;
467 do_div(dx, USEC_PER_SEC);
471 /* convert sm (bytes/psched us) into m (bps) */
477 m = (sm * PSCHED_JIFFIE2US(HZ)) >> SM_SHIFT;
481 /* convert dx (psched us) into d (us) */
487 d = dx * USEC_PER_SEC;
488 do_div(d, PSCHED_JIFFIE2US(HZ));
493 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
495 isc->sm1 = m2sm(sc->m1);
496 isc->ism1 = m2ism(sc->m1);
497 isc->dx = d2dx(sc->d);
498 isc->dy = seg_x2y(isc->dx, isc->sm1);
499 isc->sm2 = m2sm(sc->m2);
500 isc->ism2 = m2ism(sc->m2);
504 * initialize the runtime service curve with the given internal
505 * service curve starting at (x, y).
508 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
512 rtsc->sm1 = isc->sm1;
513 rtsc->ism1 = isc->ism1;
516 rtsc->sm2 = isc->sm2;
517 rtsc->ism2 = isc->ism2;
521 * calculate the y-projection of the runtime service curve by the
522 * given x-projection value
525 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
531 else if (y <= rtsc->y + rtsc->dy) {
532 /* x belongs to the 1st segment */
534 x = rtsc->x + rtsc->dx;
536 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
538 /* x belongs to the 2nd segment */
539 x = rtsc->x + rtsc->dx
540 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
546 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
552 else if (x <= rtsc->x + rtsc->dx)
553 /* y belongs to the 1st segment */
554 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
556 /* y belongs to the 2nd segment */
557 y = rtsc->y + rtsc->dy
558 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
563 * update the runtime service curve by taking the minimum of the current
564 * runtime service curve and the service curve starting at (x, y).
567 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
572 if (isc->sm1 <= isc->sm2) {
573 /* service curve is convex */
574 y1 = rtsc_x2y(rtsc, x);
576 /* the current rtsc is smaller */
584 * service curve is concave
585 * compute the two y values of the current rtsc
589 y1 = rtsc_x2y(rtsc, x);
591 /* rtsc is below isc, no change to rtsc */
595 y2 = rtsc_x2y(rtsc, x + isc->dx);
596 if (y2 >= y + isc->dy) {
597 /* rtsc is above isc, replace rtsc by isc */
606 * the two curves intersect
607 * compute the offsets (dx, dy) using the reverse
608 * function of seg_x2y()
609 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
611 dx = (y1 - y) << SM_SHIFT;
612 dsm = isc->sm1 - isc->sm2;
615 * check if (x, y1) belongs to the 1st segment of rtsc.
616 * if so, add the offset.
618 if (rtsc->x + rtsc->dx > x)
619 dx += rtsc->x + rtsc->dx - x;
620 dy = seg_x2y(dx, isc->sm1);
630 init_ed(struct hfsc_class *cl, unsigned int next_len)
634 PSCHED_GET_TIME(cur_time);
636 /* update the deadline curve */
637 rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
640 * update the eligible curve.
641 * for concave, it is equal to the deadline curve.
642 * for convex, it is a linear curve with slope m2.
644 cl->cl_eligible = cl->cl_deadline;
645 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
646 cl->cl_eligible.dx = 0;
647 cl->cl_eligible.dy = 0;
650 /* compute e and d */
651 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
652 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
658 update_ed(struct hfsc_class *cl, unsigned int next_len)
660 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
661 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
667 update_d(struct hfsc_class *cl, unsigned int next_len)
669 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
673 update_cfmin(struct hfsc_class *cl)
675 struct rb_node *n = rb_first(&cl->cf_tree);
676 struct hfsc_class *p;
682 p = rb_entry(n, struct hfsc_class, cf_node);
683 cl->cl_cfmin = p->cl_f;
687 init_vf(struct hfsc_class *cl, unsigned int len)
689 struct hfsc_class *max_cl;
696 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
697 if (go_active && cl->cl_nactive++ == 0)
703 n = rb_last(&cl->cl_parent->vt_tree);
705 max_cl = rb_entry(n, struct hfsc_class,vt_node);
707 * set vt to the average of the min and max
708 * classes. if the parent's period didn't
709 * change, don't decrease vt of the class.
712 if (cl->cl_parent->cl_cvtmin != 0)
713 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
715 if (cl->cl_parent->cl_vtperiod !=
716 cl->cl_parentperiod || vt > cl->cl_vt)
720 * first child for a new parent backlog period.
721 * add parent's cvtmax to cvtoff to make a new
722 * vt (vtoff + vt) larger than the vt in the
723 * last period for all children.
725 vt = cl->cl_parent->cl_cvtmax;
726 cl->cl_parent->cl_cvtoff += vt;
727 cl->cl_parent->cl_cvtmax = 0;
728 cl->cl_parent->cl_cvtmin = 0;
732 cl->cl_vtoff = cl->cl_parent->cl_cvtoff -
735 /* update the virtual curve */
736 vt = cl->cl_vt + cl->cl_vtoff;
737 rtsc_min(&cl->cl_virtual, &cl->cl_fsc, vt,
739 if (cl->cl_virtual.x == vt) {
740 cl->cl_virtual.x -= cl->cl_vtoff;
745 cl->cl_vtperiod++; /* increment vt period */
746 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
747 if (cl->cl_parent->cl_nactive == 0)
748 cl->cl_parentperiod++;
754 if (cl->cl_flags & HFSC_USC) {
755 /* class has upper limit curve */
757 PSCHED_GET_TIME(cur_time);
759 /* update the ulimit curve */
760 rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
763 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
769 f = max(cl->cl_myf, cl->cl_cfmin);
773 update_cfmin(cl->cl_parent);
779 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
781 u64 f; /* , myf_bound, delta; */
784 if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
787 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
790 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
793 if (go_passive && --cl->cl_nactive == 0)
799 /* no more active child, going passive */
801 /* update cvtmax of the parent class */
802 if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
803 cl->cl_parent->cl_cvtmax = cl->cl_vt;
805 /* remove this class from the vt tree */
809 update_cfmin(cl->cl_parent);
817 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
818 - cl->cl_vtoff + cl->cl_vtadj;
821 * if vt of the class is smaller than cvtmin,
822 * the class was skipped in the past due to non-fit.
823 * if so, we need to adjust vtadj.
825 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
826 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
827 cl->cl_vt = cl->cl_parent->cl_cvtmin;
830 /* update the vt tree */
833 if (cl->cl_flags & HFSC_USC) {
834 cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
838 * This code causes classes to stay way under their
839 * limit when multiple classes are used at gigabit
840 * speed. needs investigation. -kaber
843 * if myf lags behind by more than one clock tick
844 * from the current time, adjust myfadj to prevent
845 * a rate-limited class from going greedy.
846 * in a steady state under rate-limiting, myf
847 * fluctuates within one clock tick.
849 myf_bound = cur_time - PSCHED_JIFFIE2US(1);
850 if (cl->cl_myf < myf_bound) {
851 delta = cur_time - cl->cl_myf;
852 cl->cl_myfadj += delta;
858 f = max(cl->cl_myf, cl->cl_cfmin);
862 update_cfmin(cl->cl_parent);
868 set_active(struct hfsc_class *cl, unsigned int len)
870 if (cl->cl_flags & HFSC_RSC)
872 if (cl->cl_flags & HFSC_FSC)
875 list_add_tail(&cl->dlist, &cl->sched->droplist);
879 set_passive(struct hfsc_class *cl)
881 if (cl->cl_flags & HFSC_RSC)
884 list_del(&cl->dlist);
887 * vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
888 * needs to be called explicitly to remove a class from vttree.
893 * hack to get length of first packet in queue.
896 qdisc_peek_len(struct Qdisc *sch)
901 skb = sch->dequeue(sch);
904 printk("qdisc_peek_len: non work-conserving qdisc ?\n");
908 if (unlikely(sch->ops->requeue(skb, sch) != NET_XMIT_SUCCESS)) {
910 printk("qdisc_peek_len: failed to requeue\n");
911 qdisc_tree_decrease_qlen(sch, 1);
918 hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl)
920 unsigned int len = cl->qdisc->q.qlen;
922 qdisc_reset(cl->qdisc);
923 qdisc_tree_decrease_qlen(cl->qdisc, len);
927 hfsc_adjust_levels(struct hfsc_class *cl)
929 struct hfsc_class *p;
934 list_for_each_entry(p, &cl->children, siblings) {
935 if (p->level >= level)
936 level = p->level + 1;
939 } while ((cl = cl->cl_parent) != NULL);
942 static inline unsigned int
948 return h & (HFSC_HSIZE - 1);
951 static inline struct hfsc_class *
952 hfsc_find_class(u32 classid, struct Qdisc *sch)
954 struct hfsc_sched *q = qdisc_priv(sch);
955 struct hfsc_class *cl;
957 list_for_each_entry(cl, &q->clhash[hfsc_hash(classid)], hlist) {
958 if (cl->classid == classid)
965 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
968 sc2isc(rsc, &cl->cl_rsc);
969 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
970 cl->cl_eligible = cl->cl_deadline;
971 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
972 cl->cl_eligible.dx = 0;
973 cl->cl_eligible.dy = 0;
975 cl->cl_flags |= HFSC_RSC;
979 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
981 sc2isc(fsc, &cl->cl_fsc);
982 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
983 cl->cl_flags |= HFSC_FSC;
987 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
990 sc2isc(usc, &cl->cl_usc);
991 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
992 cl->cl_flags |= HFSC_USC;
996 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
997 struct rtattr **tca, unsigned long *arg)
999 struct hfsc_sched *q = qdisc_priv(sch);
1000 struct hfsc_class *cl = (struct hfsc_class *)*arg;
1001 struct hfsc_class *parent = NULL;
1002 struct rtattr *opt = tca[TCA_OPTIONS-1];
1003 struct rtattr *tb[TCA_HFSC_MAX];
1004 struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
1007 if (opt == NULL || rtattr_parse_nested(tb, TCA_HFSC_MAX, opt))
1010 if (tb[TCA_HFSC_RSC-1]) {
1011 if (RTA_PAYLOAD(tb[TCA_HFSC_RSC-1]) < sizeof(*rsc))
1013 rsc = RTA_DATA(tb[TCA_HFSC_RSC-1]);
1014 if (rsc->m1 == 0 && rsc->m2 == 0)
1018 if (tb[TCA_HFSC_FSC-1]) {
1019 if (RTA_PAYLOAD(tb[TCA_HFSC_FSC-1]) < sizeof(*fsc))
1021 fsc = RTA_DATA(tb[TCA_HFSC_FSC-1]);
1022 if (fsc->m1 == 0 && fsc->m2 == 0)
1026 if (tb[TCA_HFSC_USC-1]) {
1027 if (RTA_PAYLOAD(tb[TCA_HFSC_USC-1]) < sizeof(*usc))
1029 usc = RTA_DATA(tb[TCA_HFSC_USC-1]);
1030 if (usc->m1 == 0 && usc->m2 == 0)
1036 if (cl->cl_parent && cl->cl_parent->classid != parentid)
1038 if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
1041 PSCHED_GET_TIME(cur_time);
1045 hfsc_change_rsc(cl, rsc, cur_time);
1047 hfsc_change_fsc(cl, fsc);
1049 hfsc_change_usc(cl, usc, cur_time);
1051 if (cl->qdisc->q.qlen != 0) {
1052 if (cl->cl_flags & HFSC_RSC)
1053 update_ed(cl, qdisc_peek_len(cl->qdisc));
1054 if (cl->cl_flags & HFSC_FSC)
1055 update_vf(cl, 0, cur_time);
1057 sch_tree_unlock(sch);
1059 #ifdef CONFIG_NET_ESTIMATOR
1060 if (tca[TCA_RATE-1])
1061 gen_replace_estimator(&cl->bstats, &cl->rate_est,
1062 cl->stats_lock, tca[TCA_RATE-1]);
1067 if (parentid == TC_H_ROOT)
1072 parent = hfsc_find_class(parentid, sch);
1077 if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1079 if (hfsc_find_class(classid, sch))
1082 if (rsc == NULL && fsc == NULL)
1085 cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1090 hfsc_change_rsc(cl, rsc, 0);
1092 hfsc_change_fsc(cl, fsc);
1094 hfsc_change_usc(cl, usc, 0);
1097 cl->classid = classid;
1099 cl->cl_parent = parent;
1100 cl->qdisc = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops, classid);
1101 if (cl->qdisc == NULL)
1102 cl->qdisc = &noop_qdisc;
1103 cl->stats_lock = &sch->dev->queue_lock;
1104 INIT_LIST_HEAD(&cl->children);
1105 cl->vt_tree = RB_ROOT;
1106 cl->cf_tree = RB_ROOT;
1109 list_add_tail(&cl->hlist, &q->clhash[hfsc_hash(classid)]);
1110 list_add_tail(&cl->siblings, &parent->children);
1111 if (parent->level == 0)
1112 hfsc_purge_queue(sch, parent);
1113 hfsc_adjust_levels(parent);
1114 cl->cl_pcvtoff = parent->cl_cvtoff;
1115 sch_tree_unlock(sch);
1117 #ifdef CONFIG_NET_ESTIMATOR
1118 if (tca[TCA_RATE-1])
1119 gen_new_estimator(&cl->bstats, &cl->rate_est,
1120 cl->stats_lock, tca[TCA_RATE-1]);
1122 *arg = (unsigned long)cl;
1127 hfsc_destroy_filters(struct tcf_proto **fl)
1129 struct tcf_proto *tp;
1131 while ((tp = *fl) != NULL) {
1138 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1140 struct hfsc_sched *q = qdisc_priv(sch);
1142 hfsc_destroy_filters(&cl->filter_list);
1143 qdisc_destroy(cl->qdisc);
1144 #ifdef CONFIG_NET_ESTIMATOR
1145 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1152 hfsc_delete_class(struct Qdisc *sch, unsigned long arg)
1154 struct hfsc_sched *q = qdisc_priv(sch);
1155 struct hfsc_class *cl = (struct hfsc_class *)arg;
1157 if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1162 list_del(&cl->siblings);
1163 hfsc_adjust_levels(cl->cl_parent);
1165 hfsc_purge_queue(sch, cl);
1166 list_del(&cl->hlist);
1168 if (--cl->refcnt == 0)
1169 hfsc_destroy_class(sch, cl);
1171 sch_tree_unlock(sch);
1175 static struct hfsc_class *
1176 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1178 struct hfsc_sched *q = qdisc_priv(sch);
1179 struct hfsc_class *cl;
1180 struct tcf_result res;
1181 struct tcf_proto *tcf;
1184 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1185 (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1189 *qerr = NET_XMIT_BYPASS;
1190 tcf = q->root.filter_list;
1191 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
1192 #ifdef CONFIG_NET_CLS_ACT
1196 *qerr = NET_XMIT_SUCCESS;
1200 #elif defined(CONFIG_NET_CLS_POLICE)
1201 if (result == TC_POLICE_SHOT)
1204 if ((cl = (struct hfsc_class *)res.class) == NULL) {
1205 if ((cl = hfsc_find_class(res.classid, sch)) == NULL)
1206 break; /* filter selected invalid classid */
1210 return cl; /* hit leaf class */
1212 /* apply inner filter chain */
1213 tcf = cl->filter_list;
1216 /* classification failed, try default class */
1217 cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1218 if (cl == NULL || cl->level > 0)
1225 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1228 struct hfsc_class *cl = (struct hfsc_class *)arg;
1235 new = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops,
1242 hfsc_purge_queue(sch, cl);
1243 *old = xchg(&cl->qdisc, new);
1244 sch_tree_unlock(sch);
1248 static struct Qdisc *
1249 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1251 struct hfsc_class *cl = (struct hfsc_class *)arg;
1253 if (cl != NULL && cl->level == 0)
1260 hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1262 struct hfsc_class *cl = (struct hfsc_class *)arg;
1264 if (cl->qdisc->q.qlen == 0) {
1265 update_vf(cl, 0, 0);
1270 static unsigned long
1271 hfsc_get_class(struct Qdisc *sch, u32 classid)
1273 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1278 return (unsigned long)cl;
1282 hfsc_put_class(struct Qdisc *sch, unsigned long arg)
1284 struct hfsc_class *cl = (struct hfsc_class *)arg;
1286 if (--cl->refcnt == 0)
1287 hfsc_destroy_class(sch, cl);
1290 static unsigned long
1291 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1293 struct hfsc_class *p = (struct hfsc_class *)parent;
1294 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1297 if (p != NULL && p->level <= cl->level)
1302 return (unsigned long)cl;
1306 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1308 struct hfsc_class *cl = (struct hfsc_class *)arg;
1313 static struct tcf_proto **
1314 hfsc_tcf_chain(struct Qdisc *sch, unsigned long arg)
1316 struct hfsc_sched *q = qdisc_priv(sch);
1317 struct hfsc_class *cl = (struct hfsc_class *)arg;
1322 return &cl->filter_list;
1326 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1328 struct tc_service_curve tsc;
1330 tsc.m1 = sm2m(sc->sm1);
1331 tsc.d = dx2d(sc->dx);
1332 tsc.m2 = sm2m(sc->sm2);
1333 RTA_PUT(skb, attr, sizeof(tsc), &tsc);
1342 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1344 if ((cl->cl_flags & HFSC_RSC) &&
1345 (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1346 goto rtattr_failure;
1348 if ((cl->cl_flags & HFSC_FSC) &&
1349 (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1350 goto rtattr_failure;
1352 if ((cl->cl_flags & HFSC_USC) &&
1353 (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1354 goto rtattr_failure;
1363 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1366 struct hfsc_class *cl = (struct hfsc_class *)arg;
1367 unsigned char *b = skb->tail;
1368 struct rtattr *rta = (struct rtattr *)b;
1370 tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->classid : TC_H_ROOT;
1371 tcm->tcm_handle = cl->classid;
1373 tcm->tcm_info = cl->qdisc->handle;
1375 RTA_PUT(skb, TCA_OPTIONS, 0, NULL);
1376 if (hfsc_dump_curves(skb, cl) < 0)
1377 goto rtattr_failure;
1378 rta->rta_len = skb->tail - b;
1382 skb_trim(skb, b - skb->data);
1387 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1388 struct gnet_dump *d)
1390 struct hfsc_class *cl = (struct hfsc_class *)arg;
1391 struct tc_hfsc_stats xstats;
1393 cl->qstats.qlen = cl->qdisc->q.qlen;
1394 xstats.level = cl->level;
1395 xstats.period = cl->cl_vtperiod;
1396 xstats.work = cl->cl_total;
1397 xstats.rtwork = cl->cl_cumul;
1399 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1400 #ifdef CONFIG_NET_ESTIMATOR
1401 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1403 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1406 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1412 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1414 struct hfsc_sched *q = qdisc_priv(sch);
1415 struct hfsc_class *cl;
1421 for (i = 0; i < HFSC_HSIZE; i++) {
1422 list_for_each_entry(cl, &q->clhash[i], hlist) {
1423 if (arg->count < arg->skip) {
1427 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1437 hfsc_watchdog(unsigned long arg)
1439 struct Qdisc *sch = (struct Qdisc *)arg;
1441 sch->flags &= ~TCQ_F_THROTTLED;
1442 netif_schedule(sch->dev);
1446 hfsc_schedule_watchdog(struct Qdisc *sch, u64 cur_time)
1448 struct hfsc_sched *q = qdisc_priv(sch);
1449 struct hfsc_class *cl;
1453 if ((cl = eltree_get_minel(q)) != NULL)
1454 next_time = cl->cl_e;
1455 if (q->root.cl_cfmin != 0) {
1456 if (next_time == 0 || next_time > q->root.cl_cfmin)
1457 next_time = q->root.cl_cfmin;
1459 WARN_ON(next_time == 0);
1460 delay = next_time - cur_time;
1461 delay = PSCHED_US2JIFFIE(delay);
1463 sch->flags |= TCQ_F_THROTTLED;
1464 mod_timer(&q->wd_timer, jiffies + delay);
1468 hfsc_init_qdisc(struct Qdisc *sch, struct rtattr *opt)
1470 struct hfsc_sched *q = qdisc_priv(sch);
1471 struct tc_hfsc_qopt *qopt;
1474 if (opt == NULL || RTA_PAYLOAD(opt) < sizeof(*qopt))
1476 qopt = RTA_DATA(opt);
1478 sch->stats_lock = &sch->dev->queue_lock;
1480 q->defcls = qopt->defcls;
1481 for (i = 0; i < HFSC_HSIZE; i++)
1482 INIT_LIST_HEAD(&q->clhash[i]);
1483 q->eligible = RB_ROOT;
1484 INIT_LIST_HEAD(&q->droplist);
1485 skb_queue_head_init(&q->requeue);
1488 q->root.classid = sch->handle;
1490 q->root.qdisc = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops,
1492 if (q->root.qdisc == NULL)
1493 q->root.qdisc = &noop_qdisc;
1494 q->root.stats_lock = &sch->dev->queue_lock;
1495 INIT_LIST_HEAD(&q->root.children);
1496 q->root.vt_tree = RB_ROOT;
1497 q->root.cf_tree = RB_ROOT;
1499 list_add(&q->root.hlist, &q->clhash[hfsc_hash(q->root.classid)]);
1501 init_timer(&q->wd_timer);
1502 q->wd_timer.function = hfsc_watchdog;
1503 q->wd_timer.data = (unsigned long)sch;
1509 hfsc_change_qdisc(struct Qdisc *sch, struct rtattr *opt)
1511 struct hfsc_sched *q = qdisc_priv(sch);
1512 struct tc_hfsc_qopt *qopt;
1514 if (opt == NULL || RTA_PAYLOAD(opt) < sizeof(*qopt))
1516 qopt = RTA_DATA(opt);
1519 q->defcls = qopt->defcls;
1520 sch_tree_unlock(sch);
1526 hfsc_reset_class(struct hfsc_class *cl)
1539 cl->cl_vtperiod = 0;
1540 cl->cl_parentperiod = 0;
1547 cl->vt_tree = RB_ROOT;
1548 cl->cf_tree = RB_ROOT;
1549 qdisc_reset(cl->qdisc);
1551 if (cl->cl_flags & HFSC_RSC)
1552 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1553 if (cl->cl_flags & HFSC_FSC)
1554 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1555 if (cl->cl_flags & HFSC_USC)
1556 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1560 hfsc_reset_qdisc(struct Qdisc *sch)
1562 struct hfsc_sched *q = qdisc_priv(sch);
1563 struct hfsc_class *cl;
1566 for (i = 0; i < HFSC_HSIZE; i++) {
1567 list_for_each_entry(cl, &q->clhash[i], hlist)
1568 hfsc_reset_class(cl);
1570 __skb_queue_purge(&q->requeue);
1571 q->eligible = RB_ROOT;
1572 INIT_LIST_HEAD(&q->droplist);
1573 del_timer(&q->wd_timer);
1574 sch->flags &= ~TCQ_F_THROTTLED;
1579 hfsc_destroy_qdisc(struct Qdisc *sch)
1581 struct hfsc_sched *q = qdisc_priv(sch);
1582 struct hfsc_class *cl, *next;
1585 for (i = 0; i < HFSC_HSIZE; i++) {
1586 list_for_each_entry_safe(cl, next, &q->clhash[i], hlist)
1587 hfsc_destroy_class(sch, cl);
1589 __skb_queue_purge(&q->requeue);
1590 del_timer(&q->wd_timer);
1594 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1596 struct hfsc_sched *q = qdisc_priv(sch);
1597 unsigned char *b = skb->tail;
1598 struct tc_hfsc_qopt qopt;
1600 qopt.defcls = q->defcls;
1601 RTA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
1605 skb_trim(skb, b - skb->data);
1610 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1612 struct hfsc_class *cl;
1616 cl = hfsc_classify(skb, sch, &err);
1618 if (err == NET_XMIT_BYPASS)
1619 sch->qstats.drops++;
1625 err = cl->qdisc->enqueue(skb, cl->qdisc);
1626 if (unlikely(err != NET_XMIT_SUCCESS)) {
1628 sch->qstats.drops++;
1632 if (cl->qdisc->q.qlen == 1)
1633 set_active(cl, len);
1635 cl->bstats.packets++;
1636 cl->bstats.bytes += len;
1637 sch->bstats.packets++;
1638 sch->bstats.bytes += len;
1641 return NET_XMIT_SUCCESS;
1644 static struct sk_buff *
1645 hfsc_dequeue(struct Qdisc *sch)
1647 struct hfsc_sched *q = qdisc_priv(sch);
1648 struct hfsc_class *cl;
1649 struct sk_buff *skb;
1651 unsigned int next_len;
1654 if (sch->q.qlen == 0)
1656 if ((skb = __skb_dequeue(&q->requeue)))
1659 PSCHED_GET_TIME(cur_time);
1662 * if there are eligible classes, use real-time criteria.
1663 * find the class with the minimum deadline among
1664 * the eligible classes.
1666 if ((cl = eltree_get_mindl(q, cur_time)) != NULL) {
1670 * use link-sharing criteria
1671 * get the class with the minimum vt in the hierarchy
1673 cl = vttree_get_minvt(&q->root, cur_time);
1675 sch->qstats.overlimits++;
1676 hfsc_schedule_watchdog(sch, cur_time);
1681 skb = cl->qdisc->dequeue(cl->qdisc);
1683 if (net_ratelimit())
1684 printk("HFSC: Non-work-conserving qdisc ?\n");
1688 update_vf(cl, skb->len, cur_time);
1690 cl->cl_cumul += skb->len;
1692 if (cl->qdisc->q.qlen != 0) {
1693 if (cl->cl_flags & HFSC_RSC) {
1695 next_len = qdisc_peek_len(cl->qdisc);
1697 update_ed(cl, next_len);
1699 update_d(cl, next_len);
1702 /* the class becomes passive */
1707 sch->flags &= ~TCQ_F_THROTTLED;
1714 hfsc_requeue(struct sk_buff *skb, struct Qdisc *sch)
1716 struct hfsc_sched *q = qdisc_priv(sch);
1718 __skb_queue_head(&q->requeue, skb);
1720 sch->qstats.requeues++;
1721 return NET_XMIT_SUCCESS;
1725 hfsc_drop(struct Qdisc *sch)
1727 struct hfsc_sched *q = qdisc_priv(sch);
1728 struct hfsc_class *cl;
1731 list_for_each_entry(cl, &q->droplist, dlist) {
1732 if (cl->qdisc->ops->drop != NULL &&
1733 (len = cl->qdisc->ops->drop(cl->qdisc)) > 0) {
1734 if (cl->qdisc->q.qlen == 0) {
1735 update_vf(cl, 0, 0);
1738 list_move_tail(&cl->dlist, &q->droplist);
1741 sch->qstats.drops++;
1749 static struct Qdisc_class_ops hfsc_class_ops = {
1750 .change = hfsc_change_class,
1751 .delete = hfsc_delete_class,
1752 .graft = hfsc_graft_class,
1753 .leaf = hfsc_class_leaf,
1754 .qlen_notify = hfsc_qlen_notify,
1755 .get = hfsc_get_class,
1756 .put = hfsc_put_class,
1757 .bind_tcf = hfsc_bind_tcf,
1758 .unbind_tcf = hfsc_unbind_tcf,
1759 .tcf_chain = hfsc_tcf_chain,
1760 .dump = hfsc_dump_class,
1761 .dump_stats = hfsc_dump_class_stats,
1765 static struct Qdisc_ops hfsc_qdisc_ops = {
1767 .init = hfsc_init_qdisc,
1768 .change = hfsc_change_qdisc,
1769 .reset = hfsc_reset_qdisc,
1770 .destroy = hfsc_destroy_qdisc,
1771 .dump = hfsc_dump_qdisc,
1772 .enqueue = hfsc_enqueue,
1773 .dequeue = hfsc_dequeue,
1774 .requeue = hfsc_requeue,
1776 .cl_ops = &hfsc_class_ops,
1777 .priv_size = sizeof(struct hfsc_sched),
1778 .owner = THIS_MODULE
1784 return register_qdisc(&hfsc_qdisc_ops);
1790 unregister_qdisc(&hfsc_qdisc_ops);
1793 MODULE_LICENSE("GPL");
1794 module_init(hfsc_init);
1795 module_exit(hfsc_cleanup);