blob: e68cb440756a4c33ce1108d464ea51b1c0b95631 [file] [log] [blame]
stephen hemminger0545a302011-04-04 05:30:58 +00001/*
2 * net/sched/sch_qfq.c Quick Fair Queueing Scheduler.
3 *
4 * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente.
5 *
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * version 2 as published by the Free Software Foundation.
9 */
10
11#include <linux/module.h>
12#include <linux/init.h>
13#include <linux/bitops.h>
14#include <linux/errno.h>
15#include <linux/netdevice.h>
16#include <linux/pkt_sched.h>
17#include <net/sch_generic.h>
18#include <net/pkt_sched.h>
19#include <net/pkt_cls.h>
20
21
22/* Quick Fair Queueing
23 ===================
24
25 Sources:
26
27 Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
28 Packet Scheduling with Tight Bandwidth Distribution Guarantees."
29
30 See also:
31 http://retis.sssup.it/~fabio/linux/qfq/
32 */
33
34/*
35
36 Virtual time computations.
37
38 S, F and V are all computed in fixed point arithmetic with
39 FRAC_BITS decimal bits.
40
41 QFQ_MAX_INDEX is the maximum index allowed for a group. We need
42 one bit per index.
43 QFQ_MAX_WSHIFT is the maximum power of two supported as a weight.
44
45 The layout of the bits is as below:
46
47 [ MTU_SHIFT ][ FRAC_BITS ]
48 [ MAX_INDEX ][ MIN_SLOT_SHIFT ]
49 ^.__grp->index = 0
50 *.__grp->slot_shift
51
52 where MIN_SLOT_SHIFT is derived by difference from the others.
53
54 The max group index corresponds to Lmax/w_min, where
55 Lmax=1<<MTU_SHIFT, w_min = 1 .
56 From this, and knowing how many groups (MAX_INDEX) we want,
57 we can derive the shift corresponding to each group.
58
59 Because we often need to compute
60 F = S + len/w_i and V = V + len/wsum
61 instead of storing w_i store the value
62 inv_w = (1<<FRAC_BITS)/w_i
63 so we can do F = S + len * inv_w * wsum.
64 We use W_TOT in the formulas so we can easily move between
65 static and adaptive weight sum.
66
67 The per-scheduler-instance data contain all the data structures
68 for the scheduler: bitmaps and bucket lists.
69
70 */
71
72/*
73 * Maximum number of consecutive slots occupied by backlogged classes
74 * inside a group.
75 */
76#define QFQ_MAX_SLOTS 32
77
78/*
79 * Shifts used for class<->group mapping. We allow class weights that are
80 * in the range [1, 2^MAX_WSHIFT], and we try to map each class i to the
81 * group with the smallest index that can support the L_i / r_i configured
82 * for the class.
83 *
84 * grp->index is the index of the group; and grp->slot_shift
85 * is the shift for the corresponding (scaled) sigma_i.
86 */
87#define QFQ_MAX_INDEX 19
88#define QFQ_MAX_WSHIFT 16
89
90#define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT)
91#define QFQ_MAX_WSUM (2*QFQ_MAX_WEIGHT)
92
93#define FRAC_BITS 30 /* fixed point arithmetic */
94#define ONE_FP (1UL << FRAC_BITS)
95#define IWSUM (ONE_FP/QFQ_MAX_WSUM)
96
97#define QFQ_MTU_SHIFT 11
98#define QFQ_MIN_SLOT_SHIFT (FRAC_BITS + QFQ_MTU_SHIFT - QFQ_MAX_INDEX)
99
100/*
101 * Possible group states. These values are used as indexes for the bitmaps
102 * array of struct qfq_queue.
103 */
104enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
105
106struct qfq_group;
107
108struct qfq_class {
109 struct Qdisc_class_common common;
110
111 unsigned int refcnt;
112 unsigned int filter_cnt;
113
114 struct gnet_stats_basic_packed bstats;
115 struct gnet_stats_queue qstats;
116 struct gnet_stats_rate_est rate_est;
117 struct Qdisc *qdisc;
118
119 struct hlist_node next; /* Link for the slot list. */
120 u64 S, F; /* flow timestamps (exact) */
121
122 /* group we belong to. In principle we would need the index,
123 * which is log_2(lmax/weight), but we never reference it
124 * directly, only the group.
125 */
126 struct qfq_group *grp;
127
128 /* these are copied from the flowset. */
129 u32 inv_w; /* ONE_FP/weight */
130 u32 lmax; /* Max packet size for this flow. */
131};
132
133struct qfq_group {
134 u64 S, F; /* group timestamps (approx). */
135 unsigned int slot_shift; /* Slot shift. */
136 unsigned int index; /* Group index. */
137 unsigned int front; /* Index of the front slot. */
138 unsigned long full_slots; /* non-empty slots */
139
140 /* Array of RR lists of active classes. */
141 struct hlist_head slots[QFQ_MAX_SLOTS];
142};
143
144struct qfq_sched {
145 struct tcf_proto *filter_list;
146 struct Qdisc_class_hash clhash;
147
148 u64 V; /* Precise virtual time. */
149 u32 wsum; /* weight sum */
150
151 unsigned long bitmaps[QFQ_MAX_STATE]; /* Group bitmaps. */
152 struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
153};
154
155static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
156{
157 struct qfq_sched *q = qdisc_priv(sch);
158 struct Qdisc_class_common *clc;
159
160 clc = qdisc_class_find(&q->clhash, classid);
161 if (clc == NULL)
162 return NULL;
163 return container_of(clc, struct qfq_class, common);
164}
165
166static void qfq_purge_queue(struct qfq_class *cl)
167{
168 unsigned int len = cl->qdisc->q.qlen;
169
170 qdisc_reset(cl->qdisc);
171 qdisc_tree_decrease_qlen(cl->qdisc, len);
172}
173
174static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
175 [TCA_QFQ_WEIGHT] = { .type = NLA_U32 },
176 [TCA_QFQ_LMAX] = { .type = NLA_U32 },
177};
178
179/*
180 * Calculate a flow index, given its weight and maximum packet length.
181 * index = log_2(maxlen/weight) but we need to apply the scaling.
182 * This is used only once at flow creation.
183 */
184static int qfq_calc_index(u32 inv_w, unsigned int maxlen)
185{
186 u64 slot_size = (u64)maxlen * inv_w;
187 unsigned long size_map;
188 int index = 0;
189
190 size_map = slot_size >> QFQ_MIN_SLOT_SHIFT;
191 if (!size_map)
192 goto out;
193
194 index = __fls(size_map) + 1; /* basically a log_2 */
195 index -= !(slot_size - (1ULL << (index + QFQ_MIN_SLOT_SHIFT - 1)));
196
197 if (index < 0)
198 index = 0;
199out:
200 pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
201 (unsigned long) ONE_FP/inv_w, maxlen, index);
202
203 return index;
204}
205
206static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
207 struct nlattr **tca, unsigned long *arg)
208{
209 struct qfq_sched *q = qdisc_priv(sch);
210 struct qfq_class *cl = (struct qfq_class *)*arg;
211 struct nlattr *tb[TCA_QFQ_MAX + 1];
212 u32 weight, lmax, inv_w;
213 int i, err;
Eric Dumazetd32ae762012-01-02 11:47:50 +0000214 int delta_w;
stephen hemminger0545a302011-04-04 05:30:58 +0000215
216 if (tca[TCA_OPTIONS] == NULL) {
217 pr_notice("qfq: no options\n");
218 return -EINVAL;
219 }
220
221 err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy);
222 if (err < 0)
223 return err;
224
225 if (tb[TCA_QFQ_WEIGHT]) {
226 weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
227 if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) {
228 pr_notice("qfq: invalid weight %u\n", weight);
229 return -EINVAL;
230 }
231 } else
232 weight = 1;
233
234 inv_w = ONE_FP / weight;
235 weight = ONE_FP / inv_w;
Eric Dumazetd32ae762012-01-02 11:47:50 +0000236 delta_w = weight - (cl ? ONE_FP / cl->inv_w : 0);
237 if (q->wsum + delta_w > QFQ_MAX_WSUM) {
stephen hemminger0545a302011-04-04 05:30:58 +0000238 pr_notice("qfq: total weight out of range (%u + %u)\n",
Eric Dumazetd32ae762012-01-02 11:47:50 +0000239 delta_w, q->wsum);
stephen hemminger0545a302011-04-04 05:30:58 +0000240 return -EINVAL;
241 }
242
243 if (tb[TCA_QFQ_LMAX]) {
244 lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
245 if (!lmax || lmax > (1UL << QFQ_MTU_SHIFT)) {
246 pr_notice("qfq: invalid max length %u\n", lmax);
247 return -EINVAL;
248 }
249 } else
250 lmax = 1UL << QFQ_MTU_SHIFT;
251
252 if (cl != NULL) {
253 if (tca[TCA_RATE]) {
254 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
255 qdisc_root_sleeping_lock(sch),
256 tca[TCA_RATE]);
257 if (err)
258 return err;
259 }
260
Eric Dumazetd32ae762012-01-02 11:47:50 +0000261 if (inv_w != cl->inv_w) {
262 sch_tree_lock(sch);
263 q->wsum += delta_w;
stephen hemminger0545a302011-04-04 05:30:58 +0000264 cl->inv_w = inv_w;
Eric Dumazetd32ae762012-01-02 11:47:50 +0000265 sch_tree_unlock(sch);
stephen hemminger0545a302011-04-04 05:30:58 +0000266 }
stephen hemminger0545a302011-04-04 05:30:58 +0000267 return 0;
268 }
269
270 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
271 if (cl == NULL)
272 return -ENOBUFS;
273
274 cl->refcnt = 1;
275 cl->common.classid = classid;
276 cl->lmax = lmax;
277 cl->inv_w = inv_w;
278 i = qfq_calc_index(cl->inv_w, cl->lmax);
279
280 cl->grp = &q->groups[i];
stephen hemminger0545a302011-04-04 05:30:58 +0000281
282 cl->qdisc = qdisc_create_dflt(sch->dev_queue,
283 &pfifo_qdisc_ops, classid);
284 if (cl->qdisc == NULL)
285 cl->qdisc = &noop_qdisc;
286
287 if (tca[TCA_RATE]) {
288 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
289 qdisc_root_sleeping_lock(sch),
290 tca[TCA_RATE]);
291 if (err) {
292 qdisc_destroy(cl->qdisc);
293 kfree(cl);
294 return err;
295 }
296 }
Eric Dumazetd32ae762012-01-02 11:47:50 +0000297 q->wsum += weight;
stephen hemminger0545a302011-04-04 05:30:58 +0000298
299 sch_tree_lock(sch);
300 qdisc_class_hash_insert(&q->clhash, &cl->common);
301 sch_tree_unlock(sch);
302
303 qdisc_class_hash_grow(sch, &q->clhash);
304
305 *arg = (unsigned long)cl;
306 return 0;
307}
308
309static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
310{
311 struct qfq_sched *q = qdisc_priv(sch);
312
313 if (cl->inv_w) {
314 q->wsum -= ONE_FP / cl->inv_w;
315 cl->inv_w = 0;
316 }
317
318 gen_kill_estimator(&cl->bstats, &cl->rate_est);
319 qdisc_destroy(cl->qdisc);
320 kfree(cl);
321}
322
323static int qfq_delete_class(struct Qdisc *sch, unsigned long arg)
324{
325 struct qfq_sched *q = qdisc_priv(sch);
326 struct qfq_class *cl = (struct qfq_class *)arg;
327
328 if (cl->filter_cnt > 0)
329 return -EBUSY;
330
331 sch_tree_lock(sch);
332
333 qfq_purge_queue(cl);
334 qdisc_class_hash_remove(&q->clhash, &cl->common);
335
336 BUG_ON(--cl->refcnt == 0);
337 /*
338 * This shouldn't happen: we "hold" one cops->get() when called
339 * from tc_ctl_tclass; the destroy method is done from cops->put().
340 */
341
342 sch_tree_unlock(sch);
343 return 0;
344}
345
346static unsigned long qfq_get_class(struct Qdisc *sch, u32 classid)
347{
348 struct qfq_class *cl = qfq_find_class(sch, classid);
349
350 if (cl != NULL)
351 cl->refcnt++;
352
353 return (unsigned long)cl;
354}
355
356static void qfq_put_class(struct Qdisc *sch, unsigned long arg)
357{
358 struct qfq_class *cl = (struct qfq_class *)arg;
359
360 if (--cl->refcnt == 0)
361 qfq_destroy_class(sch, cl);
362}
363
364static struct tcf_proto **qfq_tcf_chain(struct Qdisc *sch, unsigned long cl)
365{
366 struct qfq_sched *q = qdisc_priv(sch);
367
368 if (cl)
369 return NULL;
370
371 return &q->filter_list;
372}
373
374static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
375 u32 classid)
376{
377 struct qfq_class *cl = qfq_find_class(sch, classid);
378
379 if (cl != NULL)
380 cl->filter_cnt++;
381
382 return (unsigned long)cl;
383}
384
385static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
386{
387 struct qfq_class *cl = (struct qfq_class *)arg;
388
389 cl->filter_cnt--;
390}
391
392static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
393 struct Qdisc *new, struct Qdisc **old)
394{
395 struct qfq_class *cl = (struct qfq_class *)arg;
396
397 if (new == NULL) {
398 new = qdisc_create_dflt(sch->dev_queue,
399 &pfifo_qdisc_ops, cl->common.classid);
400 if (new == NULL)
401 new = &noop_qdisc;
402 }
403
404 sch_tree_lock(sch);
405 qfq_purge_queue(cl);
406 *old = cl->qdisc;
407 cl->qdisc = new;
408 sch_tree_unlock(sch);
409 return 0;
410}
411
412static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
413{
414 struct qfq_class *cl = (struct qfq_class *)arg;
415
416 return cl->qdisc;
417}
418
419static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
420 struct sk_buff *skb, struct tcmsg *tcm)
421{
422 struct qfq_class *cl = (struct qfq_class *)arg;
423 struct nlattr *nest;
424
425 tcm->tcm_parent = TC_H_ROOT;
426 tcm->tcm_handle = cl->common.classid;
427 tcm->tcm_info = cl->qdisc->handle;
428
429 nest = nla_nest_start(skb, TCA_OPTIONS);
430 if (nest == NULL)
431 goto nla_put_failure;
432 NLA_PUT_U32(skb, TCA_QFQ_WEIGHT, ONE_FP/cl->inv_w);
433 NLA_PUT_U32(skb, TCA_QFQ_LMAX, cl->lmax);
434 return nla_nest_end(skb, nest);
435
436nla_put_failure:
437 nla_nest_cancel(skb, nest);
438 return -EMSGSIZE;
439}
440
441static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
442 struct gnet_dump *d)
443{
444 struct qfq_class *cl = (struct qfq_class *)arg;
445 struct tc_qfq_stats xstats;
446
447 memset(&xstats, 0, sizeof(xstats));
448 cl->qdisc->qstats.qlen = cl->qdisc->q.qlen;
449
450 xstats.weight = ONE_FP/cl->inv_w;
451 xstats.lmax = cl->lmax;
452
453 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
454 gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
455 gnet_stats_copy_queue(d, &cl->qdisc->qstats) < 0)
456 return -1;
457
458 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
459}
460
461static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
462{
463 struct qfq_sched *q = qdisc_priv(sch);
464 struct qfq_class *cl;
465 struct hlist_node *n;
466 unsigned int i;
467
468 if (arg->stop)
469 return;
470
471 for (i = 0; i < q->clhash.hashsize; i++) {
472 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
473 if (arg->count < arg->skip) {
474 arg->count++;
475 continue;
476 }
477 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
478 arg->stop = 1;
479 return;
480 }
481 arg->count++;
482 }
483 }
484}
485
486static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
487 int *qerr)
488{
489 struct qfq_sched *q = qdisc_priv(sch);
490 struct qfq_class *cl;
491 struct tcf_result res;
492 int result;
493
494 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
495 pr_debug("qfq_classify: found %d\n", skb->priority);
496 cl = qfq_find_class(sch, skb->priority);
497 if (cl != NULL)
498 return cl;
499 }
500
501 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
502 result = tc_classify(skb, q->filter_list, &res);
503 if (result >= 0) {
504#ifdef CONFIG_NET_CLS_ACT
505 switch (result) {
506 case TC_ACT_QUEUED:
507 case TC_ACT_STOLEN:
508 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
509 case TC_ACT_SHOT:
510 return NULL;
511 }
512#endif
513 cl = (struct qfq_class *)res.class;
514 if (cl == NULL)
515 cl = qfq_find_class(sch, res.classid);
516 return cl;
517 }
518
519 return NULL;
520}
521
522/* Generic comparison function, handling wraparound. */
523static inline int qfq_gt(u64 a, u64 b)
524{
525 return (s64)(a - b) > 0;
526}
527
528/* Round a precise timestamp to its slotted value. */
529static inline u64 qfq_round_down(u64 ts, unsigned int shift)
530{
531 return ts & ~((1ULL << shift) - 1);
532}
533
534/* return the pointer to the group with lowest index in the bitmap */
535static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
536 unsigned long bitmap)
537{
538 int index = __ffs(bitmap);
539 return &q->groups[index];
540}
541/* Calculate a mask to mimic what would be ffs_from(). */
542static inline unsigned long mask_from(unsigned long bitmap, int from)
543{
544 return bitmap & ~((1UL << from) - 1);
545}
546
547/*
548 * The state computation relies on ER=0, IR=1, EB=2, IB=3
549 * First compute eligibility comparing grp->S, q->V,
550 * then check if someone is blocking us and possibly add EB
551 */
552static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
553{
554 /* if S > V we are not eligible */
555 unsigned int state = qfq_gt(grp->S, q->V);
556 unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
557 struct qfq_group *next;
558
559 if (mask) {
560 next = qfq_ffs(q, mask);
561 if (qfq_gt(grp->F, next->F))
562 state |= EB;
563 }
564
565 return state;
566}
567
568
569/*
570 * In principle
571 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
572 * q->bitmaps[src] &= ~mask;
573 * but we should make sure that src != dst
574 */
575static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
576 int src, int dst)
577{
578 q->bitmaps[dst] |= q->bitmaps[src] & mask;
579 q->bitmaps[src] &= ~mask;
580}
581
582static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
583{
584 unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
585 struct qfq_group *next;
586
587 if (mask) {
588 next = qfq_ffs(q, mask);
589 if (!qfq_gt(next->F, old_F))
590 return;
591 }
592
593 mask = (1UL << index) - 1;
594 qfq_move_groups(q, mask, EB, ER);
595 qfq_move_groups(q, mask, IB, IR);
596}
597
598/*
599 * perhaps
600 *
601 old_V ^= q->V;
602 old_V >>= QFQ_MIN_SLOT_SHIFT;
603 if (old_V) {
604 ...
605 }
606 *
607 */
608static void qfq_make_eligible(struct qfq_sched *q, u64 old_V)
609{
610 unsigned long vslot = q->V >> QFQ_MIN_SLOT_SHIFT;
611 unsigned long old_vslot = old_V >> QFQ_MIN_SLOT_SHIFT;
612
613 if (vslot != old_vslot) {
614 unsigned long mask = (1UL << fls(vslot ^ old_vslot)) - 1;
615 qfq_move_groups(q, mask, IR, ER);
616 qfq_move_groups(q, mask, IB, EB);
617 }
618}
619
620
621/*
622 * XXX we should make sure that slot becomes less than 32.
623 * This is guaranteed by the input values.
624 * roundedS is always cl->S rounded on grp->slot_shift bits.
625 */
626static void qfq_slot_insert(struct qfq_group *grp, struct qfq_class *cl,
627 u64 roundedS)
628{
629 u64 slot = (roundedS - grp->S) >> grp->slot_shift;
630 unsigned int i = (grp->front + slot) % QFQ_MAX_SLOTS;
631
632 hlist_add_head(&cl->next, &grp->slots[i]);
633 __set_bit(slot, &grp->full_slots);
634}
635
636/* Maybe introduce hlist_first_entry?? */
637static struct qfq_class *qfq_slot_head(struct qfq_group *grp)
638{
639 return hlist_entry(grp->slots[grp->front].first,
640 struct qfq_class, next);
641}
642
643/*
644 * remove the entry from the slot
645 */
646static void qfq_front_slot_remove(struct qfq_group *grp)
647{
648 struct qfq_class *cl = qfq_slot_head(grp);
649
650 BUG_ON(!cl);
651 hlist_del(&cl->next);
652 if (hlist_empty(&grp->slots[grp->front]))
653 __clear_bit(0, &grp->full_slots);
654}
655
656/*
657 * Returns the first full queue in a group. As a side effect,
658 * adjust the bucket list so the first non-empty bucket is at
659 * position 0 in full_slots.
660 */
661static struct qfq_class *qfq_slot_scan(struct qfq_group *grp)
662{
663 unsigned int i;
664
665 pr_debug("qfq slot_scan: grp %u full %#lx\n",
666 grp->index, grp->full_slots);
667
668 if (grp->full_slots == 0)
669 return NULL;
670
671 i = __ffs(grp->full_slots); /* zero based */
672 if (i > 0) {
673 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
674 grp->full_slots >>= i;
675 }
676
677 return qfq_slot_head(grp);
678}
679
680/*
681 * adjust the bucket list. When the start time of a group decreases,
682 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
683 * move the objects. The mask of occupied slots must be shifted
684 * because we use ffs() to find the first non-empty slot.
685 * This covers decreases in the group's start time, but what about
686 * increases of the start time ?
687 * Here too we should make sure that i is less than 32
688 */
689static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
690{
691 unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
692
693 grp->full_slots <<= i;
694 grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
695}
696
697static void qfq_update_eligible(struct qfq_sched *q, u64 old_V)
698{
699 struct qfq_group *grp;
700 unsigned long ineligible;
701
702 ineligible = q->bitmaps[IR] | q->bitmaps[IB];
703 if (ineligible) {
704 if (!q->bitmaps[ER]) {
705 grp = qfq_ffs(q, ineligible);
706 if (qfq_gt(grp->S, q->V))
707 q->V = grp->S;
708 }
709 qfq_make_eligible(q, old_V);
710 }
711}
712
713/* What is length of next packet in queue (0 if queue is empty) */
714static unsigned int qdisc_peek_len(struct Qdisc *sch)
715{
716 struct sk_buff *skb;
717
718 skb = sch->ops->peek(sch);
719 return skb ? qdisc_pkt_len(skb) : 0;
720}
721
722/*
723 * Updates the class, returns true if also the group needs to be updated.
724 */
725static bool qfq_update_class(struct qfq_group *grp, struct qfq_class *cl)
726{
727 unsigned int len = qdisc_peek_len(cl->qdisc);
728
729 cl->S = cl->F;
730 if (!len)
731 qfq_front_slot_remove(grp); /* queue is empty */
732 else {
733 u64 roundedS;
734
735 cl->F = cl->S + (u64)len * cl->inv_w;
736 roundedS = qfq_round_down(cl->S, grp->slot_shift);
737 if (roundedS == grp->S)
738 return false;
739
740 qfq_front_slot_remove(grp);
741 qfq_slot_insert(grp, cl, roundedS);
742 }
743
744 return true;
745}
746
747static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
748{
749 struct qfq_sched *q = qdisc_priv(sch);
750 struct qfq_group *grp;
751 struct qfq_class *cl;
752 struct sk_buff *skb;
753 unsigned int len;
754 u64 old_V;
755
756 if (!q->bitmaps[ER])
757 return NULL;
758
759 grp = qfq_ffs(q, q->bitmaps[ER]);
760
761 cl = qfq_slot_head(grp);
762 skb = qdisc_dequeue_peeked(cl->qdisc);
763 if (!skb) {
764 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
765 return NULL;
766 }
767
768 sch->q.qlen--;
769 qdisc_bstats_update(sch, skb);
770
771 old_V = q->V;
772 len = qdisc_pkt_len(skb);
773 q->V += (u64)len * IWSUM;
774 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
775 len, (unsigned long long) cl->F, (unsigned long long) q->V);
776
777 if (qfq_update_class(grp, cl)) {
778 u64 old_F = grp->F;
779
780 cl = qfq_slot_scan(grp);
781 if (!cl)
782 __clear_bit(grp->index, &q->bitmaps[ER]);
783 else {
784 u64 roundedS = qfq_round_down(cl->S, grp->slot_shift);
785 unsigned int s;
786
787 if (grp->S == roundedS)
788 goto skip_unblock;
789 grp->S = roundedS;
790 grp->F = roundedS + (2ULL << grp->slot_shift);
791 __clear_bit(grp->index, &q->bitmaps[ER]);
792 s = qfq_calc_state(q, grp);
793 __set_bit(grp->index, &q->bitmaps[s]);
794 }
795
796 qfq_unblock_groups(q, grp->index, old_F);
797 }
798
799skip_unblock:
800 qfq_update_eligible(q, old_V);
801
802 return skb;
803}
804
805/*
806 * Assign a reasonable start time for a new flow k in group i.
807 * Admissible values for \hat(F) are multiples of \sigma_i
808 * no greater than V+\sigma_i . Larger values mean that
809 * we had a wraparound so we consider the timestamp to be stale.
810 *
811 * If F is not stale and F >= V then we set S = F.
812 * Otherwise we should assign S = V, but this may violate
813 * the ordering in ER. So, if we have groups in ER, set S to
814 * the F_j of the first group j which would be blocking us.
815 * We are guaranteed not to move S backward because
816 * otherwise our group i would still be blocked.
817 */
818static void qfq_update_start(struct qfq_sched *q, struct qfq_class *cl)
819{
820 unsigned long mask;
Eric Dumazet6bafcac2012-01-02 05:47:57 +0000821 u64 limit, roundedF;
stephen hemminger0545a302011-04-04 05:30:58 +0000822 int slot_shift = cl->grp->slot_shift;
823
824 roundedF = qfq_round_down(cl->F, slot_shift);
Eric Dumazet6bafcac2012-01-02 05:47:57 +0000825 limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
stephen hemminger0545a302011-04-04 05:30:58 +0000826
827 if (!qfq_gt(cl->F, q->V) || qfq_gt(roundedF, limit)) {
828 /* timestamp was stale */
829 mask = mask_from(q->bitmaps[ER], cl->grp->index);
830 if (mask) {
831 struct qfq_group *next = qfq_ffs(q, mask);
832 if (qfq_gt(roundedF, next->F)) {
833 cl->S = next->F;
834 return;
835 }
836 }
837 cl->S = q->V;
838 } else /* timestamp is not stale */
839 cl->S = cl->F;
840}
841
842static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
843{
844 struct qfq_sched *q = qdisc_priv(sch);
845 struct qfq_group *grp;
846 struct qfq_class *cl;
847 int err;
848 u64 roundedS;
849 int s;
850
851 cl = qfq_classify(skb, sch, &err);
852 if (cl == NULL) {
853 if (err & __NET_XMIT_BYPASS)
854 sch->qstats.drops++;
855 kfree_skb(skb);
856 return err;
857 }
858 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
859
860 err = qdisc_enqueue(skb, cl->qdisc);
861 if (unlikely(err != NET_XMIT_SUCCESS)) {
862 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
863 if (net_xmit_drop_count(err)) {
864 cl->qstats.drops++;
865 sch->qstats.drops++;
866 }
867 return err;
868 }
869
870 bstats_update(&cl->bstats, skb);
871 ++sch->q.qlen;
872
873 /* If the new skb is not the head of queue, then done here. */
874 if (cl->qdisc->q.qlen != 1)
875 return err;
876
877 /* If reach this point, queue q was idle */
878 grp = cl->grp;
879 qfq_update_start(q, cl);
880
881 /* compute new finish time and rounded start. */
882 cl->F = cl->S + (u64)qdisc_pkt_len(skb) * cl->inv_w;
883 roundedS = qfq_round_down(cl->S, grp->slot_shift);
884
885 /*
886 * insert cl in the correct bucket.
887 * If cl->S >= grp->S we don't need to adjust the
888 * bucket list and simply go to the insertion phase.
889 * Otherwise grp->S is decreasing, we must make room
890 * in the bucket list, and also recompute the group state.
891 * Finally, if there were no flows in this group and nobody
892 * was in ER make sure to adjust V.
893 */
894 if (grp->full_slots) {
895 if (!qfq_gt(grp->S, cl->S))
896 goto skip_update;
897
898 /* create a slot for this cl->S */
899 qfq_slot_rotate(grp, roundedS);
900 /* group was surely ineligible, remove */
901 __clear_bit(grp->index, &q->bitmaps[IR]);
902 __clear_bit(grp->index, &q->bitmaps[IB]);
903 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V))
904 q->V = roundedS;
905
906 grp->S = roundedS;
907 grp->F = roundedS + (2ULL << grp->slot_shift);
908 s = qfq_calc_state(q, grp);
909 __set_bit(grp->index, &q->bitmaps[s]);
910
911 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
912 s, q->bitmaps[s],
913 (unsigned long long) cl->S,
914 (unsigned long long) cl->F,
915 (unsigned long long) q->V);
916
917skip_update:
918 qfq_slot_insert(grp, cl, roundedS);
919
920 return err;
921}
922
923
924static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
925 struct qfq_class *cl)
926{
927 unsigned int i, offset;
928 u64 roundedS;
929
930 roundedS = qfq_round_down(cl->S, grp->slot_shift);
931 offset = (roundedS - grp->S) >> grp->slot_shift;
932 i = (grp->front + offset) % QFQ_MAX_SLOTS;
933
934 hlist_del(&cl->next);
935 if (hlist_empty(&grp->slots[i]))
936 __clear_bit(offset, &grp->full_slots);
937}
938
939/*
940 * called to forcibly destroy a queue.
941 * If the queue is not in the front bucket, or if it has
942 * other queues in the front bucket, we can simply remove
943 * the queue with no other side effects.
944 * Otherwise we must propagate the event up.
945 */
946static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
947{
948 struct qfq_group *grp = cl->grp;
949 unsigned long mask;
950 u64 roundedS;
951 int s;
952
953 cl->F = cl->S;
954 qfq_slot_remove(q, grp, cl);
955
956 if (!grp->full_slots) {
957 __clear_bit(grp->index, &q->bitmaps[IR]);
958 __clear_bit(grp->index, &q->bitmaps[EB]);
959 __clear_bit(grp->index, &q->bitmaps[IB]);
960
961 if (test_bit(grp->index, &q->bitmaps[ER]) &&
962 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
963 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
964 if (mask)
965 mask = ~((1UL << __fls(mask)) - 1);
966 else
967 mask = ~0UL;
968 qfq_move_groups(q, mask, EB, ER);
969 qfq_move_groups(q, mask, IB, IR);
970 }
971 __clear_bit(grp->index, &q->bitmaps[ER]);
972 } else if (hlist_empty(&grp->slots[grp->front])) {
973 cl = qfq_slot_scan(grp);
974 roundedS = qfq_round_down(cl->S, grp->slot_shift);
975 if (grp->S != roundedS) {
976 __clear_bit(grp->index, &q->bitmaps[ER]);
977 __clear_bit(grp->index, &q->bitmaps[IR]);
978 __clear_bit(grp->index, &q->bitmaps[EB]);
979 __clear_bit(grp->index, &q->bitmaps[IB]);
980 grp->S = roundedS;
981 grp->F = roundedS + (2ULL << grp->slot_shift);
982 s = qfq_calc_state(q, grp);
983 __set_bit(grp->index, &q->bitmaps[s]);
984 }
985 }
986
987 qfq_update_eligible(q, q->V);
988}
989
990static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
991{
992 struct qfq_sched *q = qdisc_priv(sch);
993 struct qfq_class *cl = (struct qfq_class *)arg;
994
995 if (cl->qdisc->q.qlen == 0)
996 qfq_deactivate_class(q, cl);
997}
998
999static unsigned int qfq_drop(struct Qdisc *sch)
1000{
1001 struct qfq_sched *q = qdisc_priv(sch);
1002 struct qfq_group *grp;
1003 unsigned int i, j, len;
1004
1005 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1006 grp = &q->groups[i];
1007 for (j = 0; j < QFQ_MAX_SLOTS; j++) {
1008 struct qfq_class *cl;
1009 struct hlist_node *n;
1010
1011 hlist_for_each_entry(cl, n, &grp->slots[j], next) {
1012
1013 if (!cl->qdisc->ops->drop)
1014 continue;
1015
1016 len = cl->qdisc->ops->drop(cl->qdisc);
1017 if (len > 0) {
1018 sch->q.qlen--;
1019 if (!cl->qdisc->q.qlen)
1020 qfq_deactivate_class(q, cl);
1021
1022 return len;
1023 }
1024 }
1025 }
1026 }
1027
1028 return 0;
1029}
1030
1031static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1032{
1033 struct qfq_sched *q = qdisc_priv(sch);
1034 struct qfq_group *grp;
1035 int i, j, err;
1036
1037 err = qdisc_class_hash_init(&q->clhash);
1038 if (err < 0)
1039 return err;
1040
1041 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1042 grp = &q->groups[i];
1043 grp->index = i;
1044 grp->slot_shift = QFQ_MTU_SHIFT + FRAC_BITS
1045 - (QFQ_MAX_INDEX - i);
1046 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1047 INIT_HLIST_HEAD(&grp->slots[j]);
1048 }
1049
1050 return 0;
1051}
1052
1053static void qfq_reset_qdisc(struct Qdisc *sch)
1054{
1055 struct qfq_sched *q = qdisc_priv(sch);
1056 struct qfq_group *grp;
1057 struct qfq_class *cl;
1058 struct hlist_node *n, *tmp;
1059 unsigned int i, j;
1060
1061 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1062 grp = &q->groups[i];
1063 for (j = 0; j < QFQ_MAX_SLOTS; j++) {
1064 hlist_for_each_entry_safe(cl, n, tmp,
1065 &grp->slots[j], next) {
1066 qfq_deactivate_class(q, cl);
1067 }
1068 }
1069 }
1070
1071 for (i = 0; i < q->clhash.hashsize; i++) {
1072 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode)
1073 qdisc_reset(cl->qdisc);
1074 }
1075 sch->q.qlen = 0;
1076}
1077
1078static void qfq_destroy_qdisc(struct Qdisc *sch)
1079{
1080 struct qfq_sched *q = qdisc_priv(sch);
1081 struct qfq_class *cl;
1082 struct hlist_node *n, *next;
1083 unsigned int i;
1084
1085 tcf_destroy_chain(&q->filter_list);
1086
1087 for (i = 0; i < q->clhash.hashsize; i++) {
1088 hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[i],
1089 common.hnode) {
1090 qfq_destroy_class(sch, cl);
1091 }
1092 }
1093 qdisc_class_hash_destroy(&q->clhash);
1094}
1095
1096static const struct Qdisc_class_ops qfq_class_ops = {
1097 .change = qfq_change_class,
1098 .delete = qfq_delete_class,
1099 .get = qfq_get_class,
1100 .put = qfq_put_class,
1101 .tcf_chain = qfq_tcf_chain,
1102 .bind_tcf = qfq_bind_tcf,
1103 .unbind_tcf = qfq_unbind_tcf,
1104 .graft = qfq_graft_class,
1105 .leaf = qfq_class_leaf,
1106 .qlen_notify = qfq_qlen_notify,
1107 .dump = qfq_dump_class,
1108 .dump_stats = qfq_dump_class_stats,
1109 .walk = qfq_walk,
1110};
1111
1112static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1113 .cl_ops = &qfq_class_ops,
1114 .id = "qfq",
1115 .priv_size = sizeof(struct qfq_sched),
1116 .enqueue = qfq_enqueue,
1117 .dequeue = qfq_dequeue,
1118 .peek = qdisc_peek_dequeued,
1119 .drop = qfq_drop,
1120 .init = qfq_init_qdisc,
1121 .reset = qfq_reset_qdisc,
1122 .destroy = qfq_destroy_qdisc,
1123 .owner = THIS_MODULE,
1124};
1125
1126static int __init qfq_init(void)
1127{
1128 return register_qdisc(&qfq_qdisc_ops);
1129}
1130
1131static void __exit qfq_exit(void)
1132{
1133 unregister_qdisc(&qfq_qdisc_ops);
1134}
1135
1136module_init(qfq_init);
1137module_exit(qfq_exit);
1138MODULE_LICENSE("GPL");