2 * net/sched/sch_qfq.c Quick Fair Queueing Plus Scheduler.
4 * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente.
5 * Copyright (c) 2012 Paolo Valente.
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * version 2 as published by the Free Software Foundation.
12 #include <linux/module.h>
13 #include <linux/init.h>
14 #include <linux/bitops.h>
15 #include <linux/errno.h>
16 #include <linux/netdevice.h>
17 #include <linux/pkt_sched.h>
18 #include <net/sch_generic.h>
19 #include <net/pkt_sched.h>
20 #include <net/pkt_cls.h>
23 /* Quick Fair Queueing Plus
24 ========================
29 "Reducing the Execution Time of Fair-Queueing Schedulers."
30 http://algo.ing.unimo.it/people/paolo/agg-sched/agg-sched.pdf
34 [2] Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
35 Packet Scheduling with Tight Bandwidth Distribution Guarantees."
38 http://retis.sssup.it/~fabio/linux/qfq/
43 QFQ+ divides classes into aggregates of at most MAX_AGG_CLASSES
44 classes. Each aggregate is timestamped with a virtual start time S
45 and a virtual finish time F, and scheduled according to its
46 timestamps. S and F are computed as a function of a system virtual
47 time function V. The classes within each aggregate are instead
50 To speed up operations, QFQ+ divides also aggregates into a limited
51 number of groups. Which group a class belongs to depends on the
52 ratio between the maximum packet length for the class and the weight
53 of the class. Groups have their own S and F. In the end, QFQ+
54 schedules groups, then aggregates within groups, then classes within
55 aggregates. See [1] and [2] for a full description.
57 Virtual time computations.
59 S, F and V are all computed in fixed point arithmetic with
60 FRAC_BITS decimal bits.
62 QFQ_MAX_INDEX is the maximum index allowed for a group. We need
64 QFQ_MAX_WSHIFT is the maximum power of two supported as a weight.
66 The layout of the bits is as below:
68 [ MTU_SHIFT ][ FRAC_BITS ]
69 [ MAX_INDEX ][ MIN_SLOT_SHIFT ]
73 where MIN_SLOT_SHIFT is derived by difference from the others.
75 The max group index corresponds to Lmax/w_min, where
76 Lmax=1<<MTU_SHIFT, w_min = 1 .
77 From this, and knowing how many groups (MAX_INDEX) we want,
78 we can derive the shift corresponding to each group.
80 Because we often need to compute
81 F = S + len/w_i and V = V + len/wsum
82 instead of storing w_i store the value
83 inv_w = (1<<FRAC_BITS)/w_i
84 so we can do F = S + len * inv_w * wsum.
85 We use W_TOT in the formulas so we can easily move between
86 static and adaptive weight sum.
88 The per-scheduler-instance data contain all the data structures
89 for the scheduler: bitmaps and bucket lists.
94 * Maximum number of consecutive slots occupied by backlogged classes
97 #define QFQ_MAX_SLOTS 32
100 * Shifts used for aggregate<->group mapping. We allow class weights that are
101 * in the range [1, 2^MAX_WSHIFT], and we try to map each aggregate i to the
102 * group with the smallest index that can support the L_i / r_i configured
103 * for the classes in the aggregate.
105 * grp->index is the index of the group; and grp->slot_shift
106 * is the shift for the corresponding (scaled) sigma_i.
108 #define QFQ_MAX_INDEX 24
109 #define QFQ_MAX_WSHIFT 10
111 #define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT) /* see qfq_slot_insert */
112 #define QFQ_MAX_WSUM (64*QFQ_MAX_WEIGHT)
114 #define FRAC_BITS 30 /* fixed point arithmetic */
115 #define ONE_FP (1UL << FRAC_BITS)
117 #define QFQ_MTU_SHIFT 16 /* to support TSO/GSO */
118 #define QFQ_MIN_LMAX 512 /* see qfq_slot_insert */
120 #define QFQ_MAX_AGG_CLASSES 8 /* max num classes per aggregate allowed */
123 * Possible group states. These values are used as indexes for the bitmaps
124 * array of struct qfq_queue.
126 enum qfq_state
{ ER
, IR
, EB
, IB
, QFQ_MAX_STATE
};
130 struct qfq_aggregate
;
133 struct Qdisc_class_common common
;
135 unsigned int filter_cnt
;
137 struct gnet_stats_basic_packed bstats
;
138 struct gnet_stats_queue qstats
;
139 struct net_rate_estimator __rcu
*rate_est
;
141 struct list_head alist
; /* Link for active-classes list. */
142 struct qfq_aggregate
*agg
; /* Parent aggregate. */
143 int deficit
; /* DRR deficit counter. */
146 struct qfq_aggregate
{
147 struct hlist_node next
; /* Link for the slot list. */
148 u64 S
, F
; /* flow timestamps (exact) */
150 /* group we belong to. In principle we would need the index,
151 * which is log_2(lmax/weight), but we never reference it
152 * directly, only the group.
154 struct qfq_group
*grp
;
156 /* these are copied from the flowset. */
157 u32 class_weight
; /* Weight of each class in this aggregate. */
158 /* Max pkt size for the classes in this aggregate, DRR quantum. */
161 u32 inv_w
; /* ONE_FP/(sum of weights of classes in aggr.). */
162 u32 budgetmax
; /* Max budget for this aggregate. */
163 u32 initial_budget
, budget
; /* Initial and current budget. */
165 int num_classes
; /* Number of classes in this aggr. */
166 struct list_head active
; /* DRR queue of active classes. */
168 struct hlist_node nonfull_next
; /* See nonfull_aggs in qfq_sched. */
172 u64 S
, F
; /* group timestamps (approx). */
173 unsigned int slot_shift
; /* Slot shift. */
174 unsigned int index
; /* Group index. */
175 unsigned int front
; /* Index of the front slot. */
176 unsigned long full_slots
; /* non-empty slots */
178 /* Array of RR lists of active aggregates. */
179 struct hlist_head slots
[QFQ_MAX_SLOTS
];
183 struct tcf_proto __rcu
*filter_list
;
184 struct tcf_block
*block
;
185 struct Qdisc_class_hash clhash
;
187 u64 oldV
, V
; /* Precise virtual times. */
188 struct qfq_aggregate
*in_serv_agg
; /* Aggregate being served. */
189 u32 wsum
; /* weight sum */
190 u32 iwsum
; /* inverse weight sum */
192 unsigned long bitmaps
[QFQ_MAX_STATE
]; /* Group bitmaps. */
193 struct qfq_group groups
[QFQ_MAX_INDEX
+ 1]; /* The groups. */
194 u32 min_slot_shift
; /* Index of the group-0 bit in the bitmaps. */
196 u32 max_agg_classes
; /* Max number of classes per aggr. */
197 struct hlist_head nonfull_aggs
; /* Aggs with room for more classes. */
201 * Possible reasons why the timestamps of an aggregate are updated
202 * enqueue: the aggregate switches from idle to active and must scheduled
204 * requeue: the aggregate finishes its budget, so it stops being served and
205 * must be rescheduled for service
207 enum update_reason
{enqueue
, requeue
};
209 static struct qfq_class
*qfq_find_class(struct Qdisc
*sch
, u32 classid
)
211 struct qfq_sched
*q
= qdisc_priv(sch
);
212 struct Qdisc_class_common
*clc
;
214 clc
= qdisc_class_find(&q
->clhash
, classid
);
217 return container_of(clc
, struct qfq_class
, common
);
220 static void qfq_purge_queue(struct qfq_class
*cl
)
222 unsigned int len
= cl
->qdisc
->q
.qlen
;
223 unsigned int backlog
= cl
->qdisc
->qstats
.backlog
;
225 qdisc_reset(cl
->qdisc
);
226 qdisc_tree_reduce_backlog(cl
->qdisc
, len
, backlog
);
229 static const struct nla_policy qfq_policy
[TCA_QFQ_MAX
+ 1] = {
230 [TCA_QFQ_WEIGHT
] = { .type
= NLA_U32
},
231 [TCA_QFQ_LMAX
] = { .type
= NLA_U32
},
235 * Calculate a flow index, given its weight and maximum packet length.
236 * index = log_2(maxlen/weight) but we need to apply the scaling.
237 * This is used only once at flow creation.
239 static int qfq_calc_index(u32 inv_w
, unsigned int maxlen
, u32 min_slot_shift
)
241 u64 slot_size
= (u64
)maxlen
* inv_w
;
242 unsigned long size_map
;
245 size_map
= slot_size
>> min_slot_shift
;
249 index
= __fls(size_map
) + 1; /* basically a log_2 */
250 index
-= !(slot_size
- (1ULL << (index
+ min_slot_shift
- 1)));
255 pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
256 (unsigned long) ONE_FP
/inv_w
, maxlen
, index
);
261 static void qfq_deactivate_agg(struct qfq_sched
*, struct qfq_aggregate
*);
262 static void qfq_activate_agg(struct qfq_sched
*, struct qfq_aggregate
*,
265 static void qfq_init_agg(struct qfq_sched
*q
, struct qfq_aggregate
*agg
,
266 u32 lmax
, u32 weight
)
268 INIT_LIST_HEAD(&agg
->active
);
269 hlist_add_head(&agg
->nonfull_next
, &q
->nonfull_aggs
);
272 agg
->class_weight
= weight
;
275 static struct qfq_aggregate
*qfq_find_agg(struct qfq_sched
*q
,
276 u32 lmax
, u32 weight
)
278 struct qfq_aggregate
*agg
;
280 hlist_for_each_entry(agg
, &q
->nonfull_aggs
, nonfull_next
)
281 if (agg
->lmax
== lmax
&& agg
->class_weight
== weight
)
288 /* Update aggregate as a function of the new number of classes. */
289 static void qfq_update_agg(struct qfq_sched
*q
, struct qfq_aggregate
*agg
,
294 if (new_num_classes
== q
->max_agg_classes
)
295 hlist_del_init(&agg
->nonfull_next
);
297 if (agg
->num_classes
> new_num_classes
&&
298 new_num_classes
== q
->max_agg_classes
- 1) /* agg no more full */
299 hlist_add_head(&agg
->nonfull_next
, &q
->nonfull_aggs
);
301 /* The next assignment may let
302 * agg->initial_budget > agg->budgetmax
303 * hold, we will take it into account in charge_actual_service().
305 agg
->budgetmax
= new_num_classes
* agg
->lmax
;
306 new_agg_weight
= agg
->class_weight
* new_num_classes
;
307 agg
->inv_w
= ONE_FP
/new_agg_weight
;
309 if (agg
->grp
== NULL
) {
310 int i
= qfq_calc_index(agg
->inv_w
, agg
->budgetmax
,
312 agg
->grp
= &q
->groups
[i
];
316 (int) agg
->class_weight
* (new_num_classes
- agg
->num_classes
);
317 q
->iwsum
= ONE_FP
/ q
->wsum
;
319 agg
->num_classes
= new_num_classes
;
322 /* Add class to aggregate. */
323 static void qfq_add_to_agg(struct qfq_sched
*q
,
324 struct qfq_aggregate
*agg
,
325 struct qfq_class
*cl
)
329 qfq_update_agg(q
, agg
, agg
->num_classes
+1);
330 if (cl
->qdisc
->q
.qlen
> 0) { /* adding an active class */
331 list_add_tail(&cl
->alist
, &agg
->active
);
332 if (list_first_entry(&agg
->active
, struct qfq_class
, alist
) ==
333 cl
&& q
->in_serv_agg
!= agg
) /* agg was inactive */
334 qfq_activate_agg(q
, agg
, enqueue
); /* schedule agg */
338 static struct qfq_aggregate
*qfq_choose_next_agg(struct qfq_sched
*);
340 static void qfq_destroy_agg(struct qfq_sched
*q
, struct qfq_aggregate
*agg
)
342 hlist_del_init(&agg
->nonfull_next
);
343 q
->wsum
-= agg
->class_weight
;
345 q
->iwsum
= ONE_FP
/ q
->wsum
;
347 if (q
->in_serv_agg
== agg
)
348 q
->in_serv_agg
= qfq_choose_next_agg(q
);
352 /* Deschedule class from within its parent aggregate. */
353 static void qfq_deactivate_class(struct qfq_sched
*q
, struct qfq_class
*cl
)
355 struct qfq_aggregate
*agg
= cl
->agg
;
358 list_del(&cl
->alist
); /* remove from RR queue of the aggregate */
359 if (list_empty(&agg
->active
)) /* agg is now inactive */
360 qfq_deactivate_agg(q
, agg
);
363 /* Remove class from its parent aggregate. */
364 static void qfq_rm_from_agg(struct qfq_sched
*q
, struct qfq_class
*cl
)
366 struct qfq_aggregate
*agg
= cl
->agg
;
369 if (agg
->num_classes
== 1) { /* agg being emptied, destroy it */
370 qfq_destroy_agg(q
, agg
);
373 qfq_update_agg(q
, agg
, agg
->num_classes
-1);
376 /* Deschedule class and remove it from its parent aggregate. */
377 static void qfq_deact_rm_from_agg(struct qfq_sched
*q
, struct qfq_class
*cl
)
379 if (cl
->qdisc
->q
.qlen
> 0) /* class is active */
380 qfq_deactivate_class(q
, cl
);
382 qfq_rm_from_agg(q
, cl
);
385 /* Move class to a new aggregate, matching the new class weight and/or lmax */
386 static int qfq_change_agg(struct Qdisc
*sch
, struct qfq_class
*cl
, u32 weight
,
389 struct qfq_sched
*q
= qdisc_priv(sch
);
390 struct qfq_aggregate
*new_agg
= qfq_find_agg(q
, lmax
, weight
);
392 if (new_agg
== NULL
) { /* create new aggregate */
393 new_agg
= kzalloc(sizeof(*new_agg
), GFP_ATOMIC
);
396 qfq_init_agg(q
, new_agg
, lmax
, weight
);
398 qfq_deact_rm_from_agg(q
, cl
);
399 qfq_add_to_agg(q
, new_agg
, cl
);
404 static int qfq_change_class(struct Qdisc
*sch
, u32 classid
, u32 parentid
,
405 struct nlattr
**tca
, unsigned long *arg
)
407 struct qfq_sched
*q
= qdisc_priv(sch
);
408 struct qfq_class
*cl
= (struct qfq_class
*)*arg
;
409 bool existing
= false;
410 struct nlattr
*tb
[TCA_QFQ_MAX
+ 1];
411 struct qfq_aggregate
*new_agg
= NULL
;
412 u32 weight
, lmax
, inv_w
;
416 if (tca
[TCA_OPTIONS
] == NULL
) {
417 pr_notice("qfq: no options\n");
421 err
= nla_parse_nested(tb
, TCA_QFQ_MAX
, tca
[TCA_OPTIONS
], qfq_policy
,
426 if (tb
[TCA_QFQ_WEIGHT
]) {
427 weight
= nla_get_u32(tb
[TCA_QFQ_WEIGHT
]);
428 if (!weight
|| weight
> (1UL << QFQ_MAX_WSHIFT
)) {
429 pr_notice("qfq: invalid weight %u\n", weight
);
435 if (tb
[TCA_QFQ_LMAX
]) {
436 lmax
= nla_get_u32(tb
[TCA_QFQ_LMAX
]);
437 if (lmax
< QFQ_MIN_LMAX
|| lmax
> (1UL << QFQ_MTU_SHIFT
)) {
438 pr_notice("qfq: invalid max length %u\n", lmax
);
442 lmax
= psched_mtu(qdisc_dev(sch
));
444 inv_w
= ONE_FP
/ weight
;
445 weight
= ONE_FP
/ inv_w
;
448 lmax
== cl
->agg
->lmax
&&
449 weight
== cl
->agg
->class_weight
)
450 return 0; /* nothing to change */
452 delta_w
= weight
- (cl
? cl
->agg
->class_weight
: 0);
454 if (q
->wsum
+ delta_w
> QFQ_MAX_WSUM
) {
455 pr_notice("qfq: total weight out of range (%d + %u)\n",
460 if (cl
!= NULL
) { /* modify existing class */
462 err
= gen_replace_estimator(&cl
->bstats
, NULL
,
465 qdisc_root_sleeping_running(sch
),
474 /* create and init new class */
475 cl
= kzalloc(sizeof(struct qfq_class
), GFP_KERNEL
);
479 cl
->common
.classid
= classid
;
482 cl
->qdisc
= qdisc_create_dflt(sch
->dev_queue
,
483 &pfifo_qdisc_ops
, classid
);
484 if (cl
->qdisc
== NULL
)
485 cl
->qdisc
= &noop_qdisc
;
488 err
= gen_new_estimator(&cl
->bstats
, NULL
,
491 qdisc_root_sleeping_running(sch
),
497 if (cl
->qdisc
!= &noop_qdisc
)
498 qdisc_hash_add(cl
->qdisc
, true);
500 qdisc_class_hash_insert(&q
->clhash
, &cl
->common
);
501 sch_tree_unlock(sch
);
503 qdisc_class_hash_grow(sch
, &q
->clhash
);
507 new_agg
= qfq_find_agg(q
, lmax
, weight
);
508 if (new_agg
== NULL
) { /* create new aggregate */
509 sch_tree_unlock(sch
);
510 new_agg
= kzalloc(sizeof(*new_agg
), GFP_KERNEL
);
511 if (new_agg
== NULL
) {
513 gen_kill_estimator(&cl
->rate_est
);
517 qfq_init_agg(q
, new_agg
, lmax
, weight
);
520 qfq_deact_rm_from_agg(q
, cl
);
521 qfq_add_to_agg(q
, new_agg
, cl
);
522 sch_tree_unlock(sch
);
524 *arg
= (unsigned long)cl
;
528 qdisc_destroy(cl
->qdisc
);
533 static void qfq_destroy_class(struct Qdisc
*sch
, struct qfq_class
*cl
)
535 struct qfq_sched
*q
= qdisc_priv(sch
);
537 qfq_rm_from_agg(q
, cl
);
538 gen_kill_estimator(&cl
->rate_est
);
539 qdisc_destroy(cl
->qdisc
);
543 static int qfq_delete_class(struct Qdisc
*sch
, unsigned long arg
)
545 struct qfq_sched
*q
= qdisc_priv(sch
);
546 struct qfq_class
*cl
= (struct qfq_class
*)arg
;
548 if (cl
->filter_cnt
> 0)
554 qdisc_class_hash_remove(&q
->clhash
, &cl
->common
);
556 sch_tree_unlock(sch
);
558 qfq_destroy_class(sch
, cl
);
562 static unsigned long qfq_search_class(struct Qdisc
*sch
, u32 classid
)
564 return (unsigned long)qfq_find_class(sch
, classid
);
567 static struct tcf_block
*qfq_tcf_block(struct Qdisc
*sch
, unsigned long cl
)
569 struct qfq_sched
*q
= qdisc_priv(sch
);
577 static unsigned long qfq_bind_tcf(struct Qdisc
*sch
, unsigned long parent
,
580 struct qfq_class
*cl
= qfq_find_class(sch
, classid
);
585 return (unsigned long)cl
;
588 static void qfq_unbind_tcf(struct Qdisc
*sch
, unsigned long arg
)
590 struct qfq_class
*cl
= (struct qfq_class
*)arg
;
595 static int qfq_graft_class(struct Qdisc
*sch
, unsigned long arg
,
596 struct Qdisc
*new, struct Qdisc
**old
)
598 struct qfq_class
*cl
= (struct qfq_class
*)arg
;
601 new = qdisc_create_dflt(sch
->dev_queue
,
602 &pfifo_qdisc_ops
, cl
->common
.classid
);
607 *old
= qdisc_replace(sch
, new, &cl
->qdisc
);
611 static struct Qdisc
*qfq_class_leaf(struct Qdisc
*sch
, unsigned long arg
)
613 struct qfq_class
*cl
= (struct qfq_class
*)arg
;
618 static int qfq_dump_class(struct Qdisc
*sch
, unsigned long arg
,
619 struct sk_buff
*skb
, struct tcmsg
*tcm
)
621 struct qfq_class
*cl
= (struct qfq_class
*)arg
;
624 tcm
->tcm_parent
= TC_H_ROOT
;
625 tcm
->tcm_handle
= cl
->common
.classid
;
626 tcm
->tcm_info
= cl
->qdisc
->handle
;
628 nest
= nla_nest_start(skb
, TCA_OPTIONS
);
630 goto nla_put_failure
;
631 if (nla_put_u32(skb
, TCA_QFQ_WEIGHT
, cl
->agg
->class_weight
) ||
632 nla_put_u32(skb
, TCA_QFQ_LMAX
, cl
->agg
->lmax
))
633 goto nla_put_failure
;
634 return nla_nest_end(skb
, nest
);
637 nla_nest_cancel(skb
, nest
);
641 static int qfq_dump_class_stats(struct Qdisc
*sch
, unsigned long arg
,
644 struct qfq_class
*cl
= (struct qfq_class
*)arg
;
645 struct tc_qfq_stats xstats
;
647 memset(&xstats
, 0, sizeof(xstats
));
649 xstats
.weight
= cl
->agg
->class_weight
;
650 xstats
.lmax
= cl
->agg
->lmax
;
652 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch
),
653 d
, NULL
, &cl
->bstats
) < 0 ||
654 gnet_stats_copy_rate_est(d
, &cl
->rate_est
) < 0 ||
655 gnet_stats_copy_queue(d
, NULL
,
656 &cl
->qdisc
->qstats
, cl
->qdisc
->q
.qlen
) < 0)
659 return gnet_stats_copy_app(d
, &xstats
, sizeof(xstats
));
662 static void qfq_walk(struct Qdisc
*sch
, struct qdisc_walker
*arg
)
664 struct qfq_sched
*q
= qdisc_priv(sch
);
665 struct qfq_class
*cl
;
671 for (i
= 0; i
< q
->clhash
.hashsize
; i
++) {
672 hlist_for_each_entry(cl
, &q
->clhash
.hash
[i
], common
.hnode
) {
673 if (arg
->count
< arg
->skip
) {
677 if (arg
->fn(sch
, (unsigned long)cl
, arg
) < 0) {
686 static struct qfq_class
*qfq_classify(struct sk_buff
*skb
, struct Qdisc
*sch
,
689 struct qfq_sched
*q
= qdisc_priv(sch
);
690 struct qfq_class
*cl
;
691 struct tcf_result res
;
692 struct tcf_proto
*fl
;
695 if (TC_H_MAJ(skb
->priority
^ sch
->handle
) == 0) {
696 pr_debug("qfq_classify: found %d\n", skb
->priority
);
697 cl
= qfq_find_class(sch
, skb
->priority
);
702 *qerr
= NET_XMIT_SUCCESS
| __NET_XMIT_BYPASS
;
703 fl
= rcu_dereference_bh(q
->filter_list
);
704 result
= tcf_classify(skb
, fl
, &res
, false);
706 #ifdef CONFIG_NET_CLS_ACT
711 *qerr
= NET_XMIT_SUCCESS
| __NET_XMIT_STOLEN
;
717 cl
= (struct qfq_class
*)res
.class;
719 cl
= qfq_find_class(sch
, res
.classid
);
726 /* Generic comparison function, handling wraparound. */
727 static inline int qfq_gt(u64 a
, u64 b
)
729 return (s64
)(a
- b
) > 0;
732 /* Round a precise timestamp to its slotted value. */
733 static inline u64
qfq_round_down(u64 ts
, unsigned int shift
)
735 return ts
& ~((1ULL << shift
) - 1);
738 /* return the pointer to the group with lowest index in the bitmap */
739 static inline struct qfq_group
*qfq_ffs(struct qfq_sched
*q
,
740 unsigned long bitmap
)
742 int index
= __ffs(bitmap
);
743 return &q
->groups
[index
];
745 /* Calculate a mask to mimic what would be ffs_from(). */
746 static inline unsigned long mask_from(unsigned long bitmap
, int from
)
748 return bitmap
& ~((1UL << from
) - 1);
752 * The state computation relies on ER=0, IR=1, EB=2, IB=3
753 * First compute eligibility comparing grp->S, q->V,
754 * then check if someone is blocking us and possibly add EB
756 static int qfq_calc_state(struct qfq_sched
*q
, const struct qfq_group
*grp
)
758 /* if S > V we are not eligible */
759 unsigned int state
= qfq_gt(grp
->S
, q
->V
);
760 unsigned long mask
= mask_from(q
->bitmaps
[ER
], grp
->index
);
761 struct qfq_group
*next
;
764 next
= qfq_ffs(q
, mask
);
765 if (qfq_gt(grp
->F
, next
->F
))
775 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
776 * q->bitmaps[src] &= ~mask;
777 * but we should make sure that src != dst
779 static inline void qfq_move_groups(struct qfq_sched
*q
, unsigned long mask
,
782 q
->bitmaps
[dst
] |= q
->bitmaps
[src
] & mask
;
783 q
->bitmaps
[src
] &= ~mask
;
786 static void qfq_unblock_groups(struct qfq_sched
*q
, int index
, u64 old_F
)
788 unsigned long mask
= mask_from(q
->bitmaps
[ER
], index
+ 1);
789 struct qfq_group
*next
;
792 next
= qfq_ffs(q
, mask
);
793 if (!qfq_gt(next
->F
, old_F
))
797 mask
= (1UL << index
) - 1;
798 qfq_move_groups(q
, mask
, EB
, ER
);
799 qfq_move_groups(q
, mask
, IB
, IR
);
806 old_V >>= q->min_slot_shift;
812 static void qfq_make_eligible(struct qfq_sched
*q
)
814 unsigned long vslot
= q
->V
>> q
->min_slot_shift
;
815 unsigned long old_vslot
= q
->oldV
>> q
->min_slot_shift
;
817 if (vslot
!= old_vslot
) {
819 int last_flip_pos
= fls(vslot
^ old_vslot
);
821 if (last_flip_pos
> 31) /* higher than the number of groups */
822 mask
= ~0UL; /* make all groups eligible */
824 mask
= (1UL << last_flip_pos
) - 1;
826 qfq_move_groups(q
, mask
, IR
, ER
);
827 qfq_move_groups(q
, mask
, IB
, EB
);
832 * The index of the slot in which the input aggregate agg is to be
833 * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
834 * and not a '-1' because the start time of the group may be moved
835 * backward by one slot after the aggregate has been inserted, and
836 * this would cause non-empty slots to be right-shifted by one
839 * QFQ+ fully satisfies this bound to the slot index if the parameters
840 * of the classes are not changed dynamically, and if QFQ+ never
841 * happens to postpone the service of agg unjustly, i.e., it never
842 * happens that the aggregate becomes backlogged and eligible, or just
843 * eligible, while an aggregate with a higher approximated finish time
844 * is being served. In particular, in this case QFQ+ guarantees that
845 * the timestamps of agg are low enough that the slot index is never
846 * higher than 2. Unfortunately, QFQ+ cannot provide the same
847 * guarantee if it happens to unjustly postpone the service of agg, or
848 * if the parameters of some class are changed.
850 * As for the first event, i.e., an out-of-order service, the
851 * upper bound to the slot index guaranteed by QFQ+ grows to
853 * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
854 * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
856 * The following function deals with this problem by backward-shifting
857 * the timestamps of agg, if needed, so as to guarantee that the slot
858 * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
859 * cause the service of other aggregates to be postponed, yet the
860 * worst-case guarantees of these aggregates are not violated. In
861 * fact, in case of no out-of-order service, the timestamps of agg
862 * would have been even lower than they are after the backward shift,
863 * because QFQ+ would have guaranteed a maximum value equal to 2 for
864 * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
865 * service is postponed because of the backward-shift would have
866 * however waited for the service of agg before being served.
868 * The other event that may cause the slot index to be higher than 2
869 * for agg is a recent change of the parameters of some class. If the
870 * weight of a class is increased or the lmax (max_pkt_size) of the
871 * class is decreased, then a new aggregate with smaller slot size
872 * than the original parent aggregate of the class may happen to be
873 * activated. The activation of this aggregate should be properly
874 * delayed to when the service of the class has finished in the ideal
875 * system tracked by QFQ+. If the activation of the aggregate is not
876 * delayed to this reference time instant, then this aggregate may be
877 * unjustly served before other aggregates waiting for service. This
878 * may cause the above bound to the slot index to be violated for some
879 * of these unlucky aggregates.
881 * Instead of delaying the activation of the new aggregate, which is
882 * quite complex, the above-discussed capping of the slot index is
883 * used to handle also the consequences of a change of the parameters
886 static void qfq_slot_insert(struct qfq_group
*grp
, struct qfq_aggregate
*agg
,
889 u64 slot
= (roundedS
- grp
->S
) >> grp
->slot_shift
;
890 unsigned int i
; /* slot index in the bucket list */
892 if (unlikely(slot
> QFQ_MAX_SLOTS
- 2)) {
893 u64 deltaS
= roundedS
- grp
->S
-
894 ((u64
)(QFQ_MAX_SLOTS
- 2)<<grp
->slot_shift
);
897 slot
= QFQ_MAX_SLOTS
- 2;
900 i
= (grp
->front
+ slot
) % QFQ_MAX_SLOTS
;
902 hlist_add_head(&agg
->next
, &grp
->slots
[i
]);
903 __set_bit(slot
, &grp
->full_slots
);
906 /* Maybe introduce hlist_first_entry?? */
907 static struct qfq_aggregate
*qfq_slot_head(struct qfq_group
*grp
)
909 return hlist_entry(grp
->slots
[grp
->front
].first
,
910 struct qfq_aggregate
, next
);
914 * remove the entry from the slot
916 static void qfq_front_slot_remove(struct qfq_group
*grp
)
918 struct qfq_aggregate
*agg
= qfq_slot_head(grp
);
921 hlist_del(&agg
->next
);
922 if (hlist_empty(&grp
->slots
[grp
->front
]))
923 __clear_bit(0, &grp
->full_slots
);
927 * Returns the first aggregate in the first non-empty bucket of the
928 * group. As a side effect, adjusts the bucket list so the first
929 * non-empty bucket is at position 0 in full_slots.
931 static struct qfq_aggregate
*qfq_slot_scan(struct qfq_group
*grp
)
935 pr_debug("qfq slot_scan: grp %u full %#lx\n",
936 grp
->index
, grp
->full_slots
);
938 if (grp
->full_slots
== 0)
941 i
= __ffs(grp
->full_slots
); /* zero based */
943 grp
->front
= (grp
->front
+ i
) % QFQ_MAX_SLOTS
;
944 grp
->full_slots
>>= i
;
947 return qfq_slot_head(grp
);
951 * adjust the bucket list. When the start time of a group decreases,
952 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
953 * move the objects. The mask of occupied slots must be shifted
954 * because we use ffs() to find the first non-empty slot.
955 * This covers decreases in the group's start time, but what about
956 * increases of the start time ?
957 * Here too we should make sure that i is less than 32
959 static void qfq_slot_rotate(struct qfq_group
*grp
, u64 roundedS
)
961 unsigned int i
= (grp
->S
- roundedS
) >> grp
->slot_shift
;
963 grp
->full_slots
<<= i
;
964 grp
->front
= (grp
->front
- i
) % QFQ_MAX_SLOTS
;
967 static void qfq_update_eligible(struct qfq_sched
*q
)
969 struct qfq_group
*grp
;
970 unsigned long ineligible
;
972 ineligible
= q
->bitmaps
[IR
] | q
->bitmaps
[IB
];
974 if (!q
->bitmaps
[ER
]) {
975 grp
= qfq_ffs(q
, ineligible
);
976 if (qfq_gt(grp
->S
, q
->V
))
979 qfq_make_eligible(q
);
983 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
984 static void agg_dequeue(struct qfq_aggregate
*agg
,
985 struct qfq_class
*cl
, unsigned int len
)
987 qdisc_dequeue_peeked(cl
->qdisc
);
989 cl
->deficit
-= (int) len
;
991 if (cl
->qdisc
->q
.qlen
== 0) /* no more packets, remove from list */
992 list_del(&cl
->alist
);
993 else if (cl
->deficit
< qdisc_pkt_len(cl
->qdisc
->ops
->peek(cl
->qdisc
))) {
994 cl
->deficit
+= agg
->lmax
;
995 list_move_tail(&cl
->alist
, &agg
->active
);
999 static inline struct sk_buff
*qfq_peek_skb(struct qfq_aggregate
*agg
,
1000 struct qfq_class
**cl
,
1003 struct sk_buff
*skb
;
1005 *cl
= list_first_entry(&agg
->active
, struct qfq_class
, alist
);
1006 skb
= (*cl
)->qdisc
->ops
->peek((*cl
)->qdisc
);
1008 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
1010 *len
= qdisc_pkt_len(skb
);
1015 /* Update F according to the actual service received by the aggregate. */
1016 static inline void charge_actual_service(struct qfq_aggregate
*agg
)
1018 /* Compute the service received by the aggregate, taking into
1019 * account that, after decreasing the number of classes in
1020 * agg, it may happen that
1021 * agg->initial_budget - agg->budget > agg->bugdetmax
1023 u32 service_received
= min(agg
->budgetmax
,
1024 agg
->initial_budget
- agg
->budget
);
1026 agg
->F
= agg
->S
+ (u64
)service_received
* agg
->inv_w
;
1029 /* Assign a reasonable start time for a new aggregate in group i.
1030 * Admissible values for \hat(F) are multiples of \sigma_i
1031 * no greater than V+\sigma_i . Larger values mean that
1032 * we had a wraparound so we consider the timestamp to be stale.
1034 * If F is not stale and F >= V then we set S = F.
1035 * Otherwise we should assign S = V, but this may violate
1036 * the ordering in EB (see [2]). So, if we have groups in ER,
1037 * set S to the F_j of the first group j which would be blocking us.
1038 * We are guaranteed not to move S backward because
1039 * otherwise our group i would still be blocked.
1041 static void qfq_update_start(struct qfq_sched
*q
, struct qfq_aggregate
*agg
)
1044 u64 limit
, roundedF
;
1045 int slot_shift
= agg
->grp
->slot_shift
;
1047 roundedF
= qfq_round_down(agg
->F
, slot_shift
);
1048 limit
= qfq_round_down(q
->V
, slot_shift
) + (1ULL << slot_shift
);
1050 if (!qfq_gt(agg
->F
, q
->V
) || qfq_gt(roundedF
, limit
)) {
1051 /* timestamp was stale */
1052 mask
= mask_from(q
->bitmaps
[ER
], agg
->grp
->index
);
1054 struct qfq_group
*next
= qfq_ffs(q
, mask
);
1055 if (qfq_gt(roundedF
, next
->F
)) {
1056 if (qfq_gt(limit
, next
->F
))
1058 else /* preserve timestamp correctness */
1064 } else /* timestamp is not stale */
1068 /* Update the timestamps of agg before scheduling/rescheduling it for
1069 * service. In particular, assign to agg->F its maximum possible
1070 * value, i.e., the virtual finish time with which the aggregate
1071 * should be labeled if it used all its budget once in service.
1074 qfq_update_agg_ts(struct qfq_sched
*q
,
1075 struct qfq_aggregate
*agg
, enum update_reason reason
)
1077 if (reason
!= requeue
)
1078 qfq_update_start(q
, agg
);
1079 else /* just charge agg for the service received */
1082 agg
->F
= agg
->S
+ (u64
)agg
->budgetmax
* agg
->inv_w
;
1085 static void qfq_schedule_agg(struct qfq_sched
*q
, struct qfq_aggregate
*agg
);
1087 static struct sk_buff
*qfq_dequeue(struct Qdisc
*sch
)
1089 struct qfq_sched
*q
= qdisc_priv(sch
);
1090 struct qfq_aggregate
*in_serv_agg
= q
->in_serv_agg
;
1091 struct qfq_class
*cl
;
1092 struct sk_buff
*skb
= NULL
;
1093 /* next-packet len, 0 means no more active classes in in-service agg */
1094 unsigned int len
= 0;
1096 if (in_serv_agg
== NULL
)
1099 if (!list_empty(&in_serv_agg
->active
))
1100 skb
= qfq_peek_skb(in_serv_agg
, &cl
, &len
);
1103 * If there are no active classes in the in-service aggregate,
1104 * or if the aggregate has not enough budget to serve its next
1105 * class, then choose the next aggregate to serve.
1107 if (len
== 0 || in_serv_agg
->budget
< len
) {
1108 charge_actual_service(in_serv_agg
);
1110 /* recharge the budget of the aggregate */
1111 in_serv_agg
->initial_budget
= in_serv_agg
->budget
=
1112 in_serv_agg
->budgetmax
;
1114 if (!list_empty(&in_serv_agg
->active
)) {
1116 * Still active: reschedule for
1117 * service. Possible optimization: if no other
1118 * aggregate is active, then there is no point
1119 * in rescheduling this aggregate, and we can
1120 * just keep it as the in-service one. This
1121 * should be however a corner case, and to
1122 * handle it, we would need to maintain an
1123 * extra num_active_aggs field.
1125 qfq_update_agg_ts(q
, in_serv_agg
, requeue
);
1126 qfq_schedule_agg(q
, in_serv_agg
);
1127 } else if (sch
->q
.qlen
== 0) { /* no aggregate to serve */
1128 q
->in_serv_agg
= NULL
;
1133 * If we get here, there are other aggregates queued:
1134 * choose the new aggregate to serve.
1136 in_serv_agg
= q
->in_serv_agg
= qfq_choose_next_agg(q
);
1137 skb
= qfq_peek_skb(in_serv_agg
, &cl
, &len
);
1142 qdisc_qstats_backlog_dec(sch
, skb
);
1144 qdisc_bstats_update(sch
, skb
);
1146 agg_dequeue(in_serv_agg
, cl
, len
);
1147 /* If lmax is lowered, through qfq_change_class, for a class
1148 * owning pending packets with larger size than the new value
1149 * of lmax, then the following condition may hold.
1151 if (unlikely(in_serv_agg
->budget
< len
))
1152 in_serv_agg
->budget
= 0;
1154 in_serv_agg
->budget
-= len
;
1156 q
->V
+= (u64
)len
* q
->iwsum
;
1157 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1158 len
, (unsigned long long) in_serv_agg
->F
,
1159 (unsigned long long) q
->V
);
1164 static struct qfq_aggregate
*qfq_choose_next_agg(struct qfq_sched
*q
)
1166 struct qfq_group
*grp
;
1167 struct qfq_aggregate
*agg
, *new_front_agg
;
1170 qfq_update_eligible(q
);
1173 if (!q
->bitmaps
[ER
])
1176 grp
= qfq_ffs(q
, q
->bitmaps
[ER
]);
1179 agg
= qfq_slot_head(grp
);
1181 /* agg starts to be served, remove it from schedule */
1182 qfq_front_slot_remove(grp
);
1184 new_front_agg
= qfq_slot_scan(grp
);
1186 if (new_front_agg
== NULL
) /* group is now inactive, remove from ER */
1187 __clear_bit(grp
->index
, &q
->bitmaps
[ER
]);
1189 u64 roundedS
= qfq_round_down(new_front_agg
->S
,
1193 if (grp
->S
== roundedS
)
1196 grp
->F
= roundedS
+ (2ULL << grp
->slot_shift
);
1197 __clear_bit(grp
->index
, &q
->bitmaps
[ER
]);
1198 s
= qfq_calc_state(q
, grp
);
1199 __set_bit(grp
->index
, &q
->bitmaps
[s
]);
1202 qfq_unblock_groups(q
, grp
->index
, old_F
);
1207 static int qfq_enqueue(struct sk_buff
*skb
, struct Qdisc
*sch
,
1208 struct sk_buff
**to_free
)
1210 struct qfq_sched
*q
= qdisc_priv(sch
);
1211 struct qfq_class
*cl
;
1212 struct qfq_aggregate
*agg
;
1215 cl
= qfq_classify(skb
, sch
, &err
);
1217 if (err
& __NET_XMIT_BYPASS
)
1218 qdisc_qstats_drop(sch
);
1219 __qdisc_drop(skb
, to_free
);
1222 pr_debug("qfq_enqueue: cl = %x\n", cl
->common
.classid
);
1224 if (unlikely(cl
->agg
->lmax
< qdisc_pkt_len(skb
))) {
1225 pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1226 cl
->agg
->lmax
, qdisc_pkt_len(skb
), cl
->common
.classid
);
1227 err
= qfq_change_agg(sch
, cl
, cl
->agg
->class_weight
,
1228 qdisc_pkt_len(skb
));
1231 return qdisc_drop(skb
, sch
, to_free
);
1235 err
= qdisc_enqueue(skb
, cl
->qdisc
, to_free
);
1236 if (unlikely(err
!= NET_XMIT_SUCCESS
)) {
1237 pr_debug("qfq_enqueue: enqueue failed %d\n", err
);
1238 if (net_xmit_drop_count(err
)) {
1240 qdisc_qstats_drop(sch
);
1245 bstats_update(&cl
->bstats
, skb
);
1246 qdisc_qstats_backlog_inc(sch
, skb
);
1250 /* if the queue was not empty, then done here */
1251 if (cl
->qdisc
->q
.qlen
!= 1) {
1252 if (unlikely(skb
== cl
->qdisc
->ops
->peek(cl
->qdisc
)) &&
1253 list_first_entry(&agg
->active
, struct qfq_class
, alist
)
1254 == cl
&& cl
->deficit
< qdisc_pkt_len(skb
))
1255 list_move_tail(&cl
->alist
, &agg
->active
);
1260 /* schedule class for service within the aggregate */
1261 cl
->deficit
= agg
->lmax
;
1262 list_add_tail(&cl
->alist
, &agg
->active
);
1264 if (list_first_entry(&agg
->active
, struct qfq_class
, alist
) != cl
||
1265 q
->in_serv_agg
== agg
)
1266 return err
; /* non-empty or in service, nothing else to do */
1268 qfq_activate_agg(q
, agg
, enqueue
);
1274 * Schedule aggregate according to its timestamps.
1276 static void qfq_schedule_agg(struct qfq_sched
*q
, struct qfq_aggregate
*agg
)
1278 struct qfq_group
*grp
= agg
->grp
;
1282 roundedS
= qfq_round_down(agg
->S
, grp
->slot_shift
);
1285 * Insert agg in the correct bucket.
1286 * If agg->S >= grp->S we don't need to adjust the
1287 * bucket list and simply go to the insertion phase.
1288 * Otherwise grp->S is decreasing, we must make room
1289 * in the bucket list, and also recompute the group state.
1290 * Finally, if there were no flows in this group and nobody
1291 * was in ER make sure to adjust V.
1293 if (grp
->full_slots
) {
1294 if (!qfq_gt(grp
->S
, agg
->S
))
1297 /* create a slot for this agg->S */
1298 qfq_slot_rotate(grp
, roundedS
);
1299 /* group was surely ineligible, remove */
1300 __clear_bit(grp
->index
, &q
->bitmaps
[IR
]);
1301 __clear_bit(grp
->index
, &q
->bitmaps
[IB
]);
1302 } else if (!q
->bitmaps
[ER
] && qfq_gt(roundedS
, q
->V
) &&
1303 q
->in_serv_agg
== NULL
)
1307 grp
->F
= roundedS
+ (2ULL << grp
->slot_shift
);
1308 s
= qfq_calc_state(q
, grp
);
1309 __set_bit(grp
->index
, &q
->bitmaps
[s
]);
1311 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1313 (unsigned long long) agg
->S
,
1314 (unsigned long long) agg
->F
,
1315 (unsigned long long) q
->V
);
1318 qfq_slot_insert(grp
, agg
, roundedS
);
1322 /* Update agg ts and schedule agg for service */
1323 static void qfq_activate_agg(struct qfq_sched
*q
, struct qfq_aggregate
*agg
,
1324 enum update_reason reason
)
1326 agg
->initial_budget
= agg
->budget
= agg
->budgetmax
; /* recharge budg. */
1328 qfq_update_agg_ts(q
, agg
, reason
);
1329 if (q
->in_serv_agg
== NULL
) { /* no aggr. in service or scheduled */
1330 q
->in_serv_agg
= agg
; /* start serving this aggregate */
1331 /* update V: to be in service, agg must be eligible */
1332 q
->oldV
= q
->V
= agg
->S
;
1333 } else if (agg
!= q
->in_serv_agg
)
1334 qfq_schedule_agg(q
, agg
);
1337 static void qfq_slot_remove(struct qfq_sched
*q
, struct qfq_group
*grp
,
1338 struct qfq_aggregate
*agg
)
1340 unsigned int i
, offset
;
1343 roundedS
= qfq_round_down(agg
->S
, grp
->slot_shift
);
1344 offset
= (roundedS
- grp
->S
) >> grp
->slot_shift
;
1346 i
= (grp
->front
+ offset
) % QFQ_MAX_SLOTS
;
1348 hlist_del(&agg
->next
);
1349 if (hlist_empty(&grp
->slots
[i
]))
1350 __clear_bit(offset
, &grp
->full_slots
);
1354 * Called to forcibly deschedule an aggregate. If the aggregate is
1355 * not in the front bucket, or if the latter has other aggregates in
1356 * the front bucket, we can simply remove the aggregate with no other
1358 * Otherwise we must propagate the event up.
1360 static void qfq_deactivate_agg(struct qfq_sched
*q
, struct qfq_aggregate
*agg
)
1362 struct qfq_group
*grp
= agg
->grp
;
1367 if (agg
== q
->in_serv_agg
) {
1368 charge_actual_service(agg
);
1369 q
->in_serv_agg
= qfq_choose_next_agg(q
);
1374 qfq_slot_remove(q
, grp
, agg
);
1376 if (!grp
->full_slots
) {
1377 __clear_bit(grp
->index
, &q
->bitmaps
[IR
]);
1378 __clear_bit(grp
->index
, &q
->bitmaps
[EB
]);
1379 __clear_bit(grp
->index
, &q
->bitmaps
[IB
]);
1381 if (test_bit(grp
->index
, &q
->bitmaps
[ER
]) &&
1382 !(q
->bitmaps
[ER
] & ~((1UL << grp
->index
) - 1))) {
1383 mask
= q
->bitmaps
[ER
] & ((1UL << grp
->index
) - 1);
1385 mask
= ~((1UL << __fls(mask
)) - 1);
1388 qfq_move_groups(q
, mask
, EB
, ER
);
1389 qfq_move_groups(q
, mask
, IB
, IR
);
1391 __clear_bit(grp
->index
, &q
->bitmaps
[ER
]);
1392 } else if (hlist_empty(&grp
->slots
[grp
->front
])) {
1393 agg
= qfq_slot_scan(grp
);
1394 roundedS
= qfq_round_down(agg
->S
, grp
->slot_shift
);
1395 if (grp
->S
!= roundedS
) {
1396 __clear_bit(grp
->index
, &q
->bitmaps
[ER
]);
1397 __clear_bit(grp
->index
, &q
->bitmaps
[IR
]);
1398 __clear_bit(grp
->index
, &q
->bitmaps
[EB
]);
1399 __clear_bit(grp
->index
, &q
->bitmaps
[IB
]);
1401 grp
->F
= roundedS
+ (2ULL << grp
->slot_shift
);
1402 s
= qfq_calc_state(q
, grp
);
1403 __set_bit(grp
->index
, &q
->bitmaps
[s
]);
1408 static void qfq_qlen_notify(struct Qdisc
*sch
, unsigned long arg
)
1410 struct qfq_sched
*q
= qdisc_priv(sch
);
1411 struct qfq_class
*cl
= (struct qfq_class
*)arg
;
1413 qfq_deactivate_class(q
, cl
);
1416 static int qfq_init_qdisc(struct Qdisc
*sch
, struct nlattr
*opt
)
1418 struct qfq_sched
*q
= qdisc_priv(sch
);
1419 struct qfq_group
*grp
;
1421 u32 max_cl_shift
, maxbudg_shift
, max_classes
;
1423 err
= tcf_block_get(&q
->block
, &q
->filter_list
, sch
);
1427 err
= qdisc_class_hash_init(&q
->clhash
);
1431 if (qdisc_dev(sch
)->tx_queue_len
+ 1 > QFQ_MAX_AGG_CLASSES
)
1432 max_classes
= QFQ_MAX_AGG_CLASSES
;
1434 max_classes
= qdisc_dev(sch
)->tx_queue_len
+ 1;
1435 /* max_cl_shift = floor(log_2(max_classes)) */
1436 max_cl_shift
= __fls(max_classes
);
1437 q
->max_agg_classes
= 1<<max_cl_shift
;
1439 /* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1440 maxbudg_shift
= QFQ_MTU_SHIFT
+ max_cl_shift
;
1441 q
->min_slot_shift
= FRAC_BITS
+ maxbudg_shift
- QFQ_MAX_INDEX
;
1443 for (i
= 0; i
<= QFQ_MAX_INDEX
; i
++) {
1444 grp
= &q
->groups
[i
];
1446 grp
->slot_shift
= q
->min_slot_shift
+ i
;
1447 for (j
= 0; j
< QFQ_MAX_SLOTS
; j
++)
1448 INIT_HLIST_HEAD(&grp
->slots
[j
]);
1451 INIT_HLIST_HEAD(&q
->nonfull_aggs
);
1456 static void qfq_reset_qdisc(struct Qdisc
*sch
)
1458 struct qfq_sched
*q
= qdisc_priv(sch
);
1459 struct qfq_class
*cl
;
1462 for (i
= 0; i
< q
->clhash
.hashsize
; i
++) {
1463 hlist_for_each_entry(cl
, &q
->clhash
.hash
[i
], common
.hnode
) {
1464 if (cl
->qdisc
->q
.qlen
> 0)
1465 qfq_deactivate_class(q
, cl
);
1467 qdisc_reset(cl
->qdisc
);
1470 sch
->qstats
.backlog
= 0;
1474 static void qfq_destroy_qdisc(struct Qdisc
*sch
)
1476 struct qfq_sched
*q
= qdisc_priv(sch
);
1477 struct qfq_class
*cl
;
1478 struct hlist_node
*next
;
1481 tcf_block_put(q
->block
);
1483 for (i
= 0; i
< q
->clhash
.hashsize
; i
++) {
1484 hlist_for_each_entry_safe(cl
, next
, &q
->clhash
.hash
[i
],
1486 qfq_destroy_class(sch
, cl
);
1489 qdisc_class_hash_destroy(&q
->clhash
);
1492 static const struct Qdisc_class_ops qfq_class_ops
= {
1493 .change
= qfq_change_class
,
1494 .delete = qfq_delete_class
,
1495 .find
= qfq_search_class
,
1496 .tcf_block
= qfq_tcf_block
,
1497 .bind_tcf
= qfq_bind_tcf
,
1498 .unbind_tcf
= qfq_unbind_tcf
,
1499 .graft
= qfq_graft_class
,
1500 .leaf
= qfq_class_leaf
,
1501 .qlen_notify
= qfq_qlen_notify
,
1502 .dump
= qfq_dump_class
,
1503 .dump_stats
= qfq_dump_class_stats
,
1507 static struct Qdisc_ops qfq_qdisc_ops __read_mostly
= {
1508 .cl_ops
= &qfq_class_ops
,
1510 .priv_size
= sizeof(struct qfq_sched
),
1511 .enqueue
= qfq_enqueue
,
1512 .dequeue
= qfq_dequeue
,
1513 .peek
= qdisc_peek_dequeued
,
1514 .init
= qfq_init_qdisc
,
1515 .reset
= qfq_reset_qdisc
,
1516 .destroy
= qfq_destroy_qdisc
,
1517 .owner
= THIS_MODULE
,
1520 static int __init
qfq_init(void)
1522 return register_qdisc(&qfq_qdisc_ops
);
1525 static void __exit
qfq_exit(void)
1527 unregister_qdisc(&qfq_qdisc_ops
);
1530 module_init(qfq_init
);
1531 module_exit(qfq_exit
);
1532 MODULE_LICENSE("GPL");