~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/net/sched/sch_qfq.c

Version: ~ [ linux-5.6-rc3 ] ~ [ linux-5.5.6 ] ~ [ linux-5.4.22 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.106 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.171 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.214 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.214 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.82 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * net/sched/sch_qfq.c         Quick Fair Queueing Plus Scheduler.
  3  *
  4  * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente.
  5  * Copyright (c) 2012 Paolo Valente.
  6  *
  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.
 10  */
 11 
 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>
 21 
 22 
 23 /*  Quick Fair Queueing Plus
 24     ========================
 25 
 26     Sources:
 27 
 28     [1] Paolo Valente,
 29     "Reducing the Execution Time of Fair-Queueing Schedulers."
 30     http://algo.ing.unimo.it/people/paolo/agg-sched/agg-sched.pdf
 31 
 32     Sources for QFQ:
 33 
 34     [2] Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
 35     Packet Scheduling with Tight Bandwidth Distribution Guarantees."
 36 
 37     See also:
 38     http://retis.sssup.it/~fabio/linux/qfq/
 39  */
 40 
 41 /*
 42 
 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
 48   scheduled with DRR.
 49 
 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.
 56 
 57   Virtual time computations.
 58 
 59   S, F and V are all computed in fixed point arithmetic with
 60   FRAC_BITS decimal bits.
 61 
 62   QFQ_MAX_INDEX is the maximum index allowed for a group. We need
 63         one bit per index.
 64   QFQ_MAX_WSHIFT is the maximum power of two supported as a weight.
 65 
 66   The layout of the bits is as below:
 67 
 68                    [ MTU_SHIFT ][      FRAC_BITS    ]
 69                    [ MAX_INDEX    ][ MIN_SLOT_SHIFT ]
 70                                  ^.__grp->index = 0
 71                                  *.__grp->slot_shift
 72 
 73   where MIN_SLOT_SHIFT is derived by difference from the others.
 74 
 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.
 79 
 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.
 87 
 88   The per-scheduler-instance data contain all the data structures
 89   for the scheduler: bitmaps and bucket lists.
 90 
 91  */
 92 
 93 /*
 94  * Maximum number of consecutive slots occupied by backlogged classes
 95  * inside a group.
 96  */
 97 #define QFQ_MAX_SLOTS   32
 98 
 99 /*
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.
104  *
105  * grp->index is the index of the group; and grp->slot_shift
106  * is the shift for the corresponding (scaled) sigma_i.
107  */
108 #define QFQ_MAX_INDEX           24
109 #define QFQ_MAX_WSHIFT          10
110 
111 #define QFQ_MAX_WEIGHT          (1<<QFQ_MAX_WSHIFT) /* see qfq_slot_insert */
112 #define QFQ_MAX_WSUM            (64*QFQ_MAX_WEIGHT)
113 
114 #define FRAC_BITS               30      /* fixed point arithmetic */
115 #define ONE_FP                  (1UL << FRAC_BITS)
116 
117 #define QFQ_MTU_SHIFT           16      /* to support TSO/GSO */
118 #define QFQ_MIN_LMAX            512     /* see qfq_slot_insert */
119 
120 #define QFQ_MAX_AGG_CLASSES     8 /* max num classes per aggregate allowed */
121 
122 /*
123  * Possible group states.  These values are used as indexes for the bitmaps
124  * array of struct qfq_queue.
125  */
126 enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
127 
128 struct qfq_group;
129 
130 struct qfq_aggregate;
131 
132 struct qfq_class {
133         struct Qdisc_class_common common;
134 
135         unsigned int refcnt;
136         unsigned int filter_cnt;
137 
138         struct gnet_stats_basic_packed bstats;
139         struct gnet_stats_queue qstats;
140         struct gnet_stats_rate_est64 rate_est;
141         struct Qdisc *qdisc;
142         struct list_head alist;         /* Link for active-classes list. */
143         struct qfq_aggregate *agg;      /* Parent aggregate. */
144         int deficit;                    /* DRR deficit counter. */
145 };
146 
147 struct qfq_aggregate {
148         struct hlist_node next; /* Link for the slot list. */
149         u64 S, F;               /* flow timestamps (exact) */
150 
151         /* group we belong to. In principle we would need the index,
152          * which is log_2(lmax/weight), but we never reference it
153          * directly, only the group.
154          */
155         struct qfq_group *grp;
156 
157         /* these are copied from the flowset. */
158         u32     class_weight; /* Weight of each class in this aggregate. */
159         /* Max pkt size for the classes in this aggregate, DRR quantum. */
160         int     lmax;
161 
162         u32     inv_w;      /* ONE_FP/(sum of weights of classes in aggr.). */
163         u32     budgetmax;  /* Max budget for this aggregate. */
164         u32     initial_budget, budget;     /* Initial and current budget. */
165 
166         int               num_classes;  /* Number of classes in this aggr. */
167         struct list_head  active;       /* DRR queue of active classes. */
168 
169         struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */
170 };
171 
172 struct qfq_group {
173         u64 S, F;                       /* group timestamps (approx). */
174         unsigned int slot_shift;        /* Slot shift. */
175         unsigned int index;             /* Group index. */
176         unsigned int front;             /* Index of the front slot. */
177         unsigned long full_slots;       /* non-empty slots */
178 
179         /* Array of RR lists of active aggregates. */
180         struct hlist_head slots[QFQ_MAX_SLOTS];
181 };
182 
183 struct qfq_sched {
184         struct tcf_proto *filter_list;
185         struct Qdisc_class_hash clhash;
186 
187         u64                     oldV, V;        /* Precise virtual times. */
188         struct qfq_aggregate    *in_serv_agg;   /* Aggregate being served. */
189         u32                     num_active_agg; /* Num. of active aggregates */
190         u32                     wsum;           /* weight sum */
191         u32                     iwsum;          /* inverse weight sum */
192 
193         unsigned long bitmaps[QFQ_MAX_STATE];       /* Group bitmaps. */
194         struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
195         u32 min_slot_shift;     /* Index of the group-0 bit in the bitmaps. */
196 
197         u32 max_agg_classes;            /* Max number of classes per aggr. */
198         struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */
199 };
200 
201 /*
202  * Possible reasons why the timestamps of an aggregate are updated
203  * enqueue: the aggregate switches from idle to active and must scheduled
204  *          for service
205  * requeue: the aggregate finishes its budget, so it stops being served and
206  *          must be rescheduled for service
207  */
208 enum update_reason {enqueue, requeue};
209 
210 static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
211 {
212         struct qfq_sched *q = qdisc_priv(sch);
213         struct Qdisc_class_common *clc;
214 
215         clc = qdisc_class_find(&q->clhash, classid);
216         if (clc == NULL)
217                 return NULL;
218         return container_of(clc, struct qfq_class, common);
219 }
220 
221 static void qfq_purge_queue(struct qfq_class *cl)
222 {
223         unsigned int len = cl->qdisc->q.qlen;
224 
225         qdisc_reset(cl->qdisc);
226         qdisc_tree_decrease_qlen(cl->qdisc, len);
227 }
228 
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 },
232 };
233 
234 /*
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.
238  */
239 static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
240 {
241         u64 slot_size = (u64)maxlen * inv_w;
242         unsigned long size_map;
243         int index = 0;
244 
245         size_map = slot_size >> min_slot_shift;
246         if (!size_map)
247                 goto out;
248 
249         index = __fls(size_map) + 1;    /* basically a log_2 */
250         index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
251 
252         if (index < 0)
253                 index = 0;
254 out:
255         pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
256                  (unsigned long) ONE_FP/inv_w, maxlen, index);
257 
258         return index;
259 }
260 
261 static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
262 static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
263                              enum update_reason);
264 
265 static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
266                          u32 lmax, u32 weight)
267 {
268         INIT_LIST_HEAD(&agg->active);
269         hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
270 
271         agg->lmax = lmax;
272         agg->class_weight = weight;
273 }
274 
275 static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
276                                           u32 lmax, u32 weight)
277 {
278         struct qfq_aggregate *agg;
279 
280         hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
281                 if (agg->lmax == lmax && agg->class_weight == weight)
282                         return agg;
283 
284         return NULL;
285 }
286 
287 
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,
290                            int new_num_classes)
291 {
292         u32 new_agg_weight;
293 
294         if (new_num_classes == q->max_agg_classes)
295                 hlist_del_init(&agg->nonfull_next);
296 
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);
300 
301         /* The next assignment may let
302          * agg->initial_budget > agg->budgetmax
303          * hold, we will take it into account in charge_actual_service().
304          */
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;
308 
309         if (agg->grp == NULL) {
310                 int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
311                                        q->min_slot_shift);
312                 agg->grp = &q->groups[i];
313         }
314 
315         q->wsum +=
316                 (int) agg->class_weight * (new_num_classes - agg->num_classes);
317         q->iwsum = ONE_FP / q->wsum;
318 
319         agg->num_classes = new_num_classes;
320 }
321 
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)
326 {
327         cl->agg = agg;
328 
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 */
335         }
336 }
337 
338 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
339 
340 static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
341 {
342         if (!hlist_unhashed(&agg->nonfull_next))
343                 hlist_del_init(&agg->nonfull_next);
344         q->wsum -= agg->class_weight;
345         if (q->wsum != 0)
346                 q->iwsum = ONE_FP / q->wsum;
347 
348         if (q->in_serv_agg == agg)
349                 q->in_serv_agg = qfq_choose_next_agg(q);
350         kfree(agg);
351 }
352 
353 /* Deschedule class from within its parent aggregate. */
354 static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
355 {
356         struct qfq_aggregate *agg = cl->agg;
357 
358 
359         list_del(&cl->alist); /* remove from RR queue of the aggregate */
360         if (list_empty(&agg->active)) /* agg is now inactive */
361                 qfq_deactivate_agg(q, agg);
362 }
363 
364 /* Remove class from its parent aggregate. */
365 static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
366 {
367         struct qfq_aggregate *agg = cl->agg;
368 
369         cl->agg = NULL;
370         if (agg->num_classes == 1) { /* agg being emptied, destroy it */
371                 qfq_destroy_agg(q, agg);
372                 return;
373         }
374         qfq_update_agg(q, agg, agg->num_classes-1);
375 }
376 
377 /* Deschedule class and remove it from its parent aggregate. */
378 static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
379 {
380         if (cl->qdisc->q.qlen > 0) /* class is active */
381                 qfq_deactivate_class(q, cl);
382 
383         qfq_rm_from_agg(q, cl);
384 }
385 
386 /* Move class to a new aggregate, matching the new class weight and/or lmax */
387 static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
388                            u32 lmax)
389 {
390         struct qfq_sched *q = qdisc_priv(sch);
391         struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
392 
393         if (new_agg == NULL) { /* create new aggregate */
394                 new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
395                 if (new_agg == NULL)
396                         return -ENOBUFS;
397                 qfq_init_agg(q, new_agg, lmax, weight);
398         }
399         qfq_deact_rm_from_agg(q, cl);
400         qfq_add_to_agg(q, new_agg, cl);
401 
402         return 0;
403 }
404 
405 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
406                             struct nlattr **tca, unsigned long *arg)
407 {
408         struct qfq_sched *q = qdisc_priv(sch);
409         struct qfq_class *cl = (struct qfq_class *)*arg;
410         bool existing = false;
411         struct nlattr *tb[TCA_QFQ_MAX + 1];
412         struct qfq_aggregate *new_agg = NULL;
413         u32 weight, lmax, inv_w;
414         int err;
415         int delta_w;
416 
417         if (tca[TCA_OPTIONS] == NULL) {
418                 pr_notice("qfq: no options\n");
419                 return -EINVAL;
420         }
421 
422         err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy);
423         if (err < 0)
424                 return err;
425 
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);
430                         return -EINVAL;
431                 }
432         } else
433                 weight = 1;
434 
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);
439                         return -EINVAL;
440                 }
441         } else
442                 lmax = psched_mtu(qdisc_dev(sch));
443 
444         inv_w = ONE_FP / weight;
445         weight = ONE_FP / inv_w;
446 
447         if (cl != NULL &&
448             lmax == cl->agg->lmax &&
449             weight == cl->agg->class_weight)
450                 return 0; /* nothing to change */
451 
452         delta_w = weight - (cl ? cl->agg->class_weight : 0);
453 
454         if (q->wsum + delta_w > QFQ_MAX_WSUM) {
455                 pr_notice("qfq: total weight out of range (%d + %u)\n",
456                           delta_w, q->wsum);
457                 return -EINVAL;
458         }
459 
460         if (cl != NULL) { /* modify existing class */
461                 if (tca[TCA_RATE]) {
462                         err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
463                                                     qdisc_root_sleeping_lock(sch),
464                                                     tca[TCA_RATE]);
465                         if (err)
466                                 return err;
467                 }
468                 existing = true;
469                 goto set_change_agg;
470         }
471 
472         /* create and init new class */
473         cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
474         if (cl == NULL)
475                 return -ENOBUFS;
476 
477         cl->refcnt = 1;
478         cl->common.classid = classid;
479         cl->deficit = lmax;
480 
481         cl->qdisc = qdisc_create_dflt(sch->dev_queue,
482                                       &pfifo_qdisc_ops, classid);
483         if (cl->qdisc == NULL)
484                 cl->qdisc = &noop_qdisc;
485 
486         if (tca[TCA_RATE]) {
487                 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
488                                         qdisc_root_sleeping_lock(sch),
489                                         tca[TCA_RATE]);
490                 if (err)
491                         goto destroy_class;
492         }
493 
494         sch_tree_lock(sch);
495         qdisc_class_hash_insert(&q->clhash, &cl->common);
496         sch_tree_unlock(sch);
497 
498         qdisc_class_hash_grow(sch, &q->clhash);
499 
500 set_change_agg:
501         sch_tree_lock(sch);
502         new_agg = qfq_find_agg(q, lmax, weight);
503         if (new_agg == NULL) { /* create new aggregate */
504                 sch_tree_unlock(sch);
505                 new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
506                 if (new_agg == NULL) {
507                         err = -ENOBUFS;
508                         gen_kill_estimator(&cl->bstats, &cl->rate_est);
509                         goto destroy_class;
510                 }
511                 sch_tree_lock(sch);
512                 qfq_init_agg(q, new_agg, lmax, weight);
513         }
514         if (existing)
515                 qfq_deact_rm_from_agg(q, cl);
516         qfq_add_to_agg(q, new_agg, cl);
517         sch_tree_unlock(sch);
518 
519         *arg = (unsigned long)cl;
520         return 0;
521 
522 destroy_class:
523         qdisc_destroy(cl->qdisc);
524         kfree(cl);
525         return err;
526 }
527 
528 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
529 {
530         struct qfq_sched *q = qdisc_priv(sch);
531 
532         qfq_rm_from_agg(q, cl);
533         gen_kill_estimator(&cl->bstats, &cl->rate_est);
534         qdisc_destroy(cl->qdisc);
535         kfree(cl);
536 }
537 
538 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg)
539 {
540         struct qfq_sched *q = qdisc_priv(sch);
541         struct qfq_class *cl = (struct qfq_class *)arg;
542 
543         if (cl->filter_cnt > 0)
544                 return -EBUSY;
545 
546         sch_tree_lock(sch);
547 
548         qfq_purge_queue(cl);
549         qdisc_class_hash_remove(&q->clhash, &cl->common);
550 
551         BUG_ON(--cl->refcnt == 0);
552         /*
553          * This shouldn't happen: we "hold" one cops->get() when called
554          * from tc_ctl_tclass; the destroy method is done from cops->put().
555          */
556 
557         sch_tree_unlock(sch);
558         return 0;
559 }
560 
561 static unsigned long qfq_get_class(struct Qdisc *sch, u32 classid)
562 {
563         struct qfq_class *cl = qfq_find_class(sch, classid);
564 
565         if (cl != NULL)
566                 cl->refcnt++;
567 
568         return (unsigned long)cl;
569 }
570 
571 static void qfq_put_class(struct Qdisc *sch, unsigned long arg)
572 {
573         struct qfq_class *cl = (struct qfq_class *)arg;
574 
575         if (--cl->refcnt == 0)
576                 qfq_destroy_class(sch, cl);
577 }
578 
579 static struct tcf_proto **qfq_tcf_chain(struct Qdisc *sch, unsigned long cl)
580 {
581         struct qfq_sched *q = qdisc_priv(sch);
582 
583         if (cl)
584                 return NULL;
585 
586         return &q->filter_list;
587 }
588 
589 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
590                                   u32 classid)
591 {
592         struct qfq_class *cl = qfq_find_class(sch, classid);
593 
594         if (cl != NULL)
595                 cl->filter_cnt++;
596 
597         return (unsigned long)cl;
598 }
599 
600 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
601 {
602         struct qfq_class *cl = (struct qfq_class *)arg;
603 
604         cl->filter_cnt--;
605 }
606 
607 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
608                            struct Qdisc *new, struct Qdisc **old)
609 {
610         struct qfq_class *cl = (struct qfq_class *)arg;
611 
612         if (new == NULL) {
613                 new = qdisc_create_dflt(sch->dev_queue,
614                                         &pfifo_qdisc_ops, cl->common.classid);
615                 if (new == NULL)
616                         new = &noop_qdisc;
617         }
618 
619         sch_tree_lock(sch);
620         qfq_purge_queue(cl);
621         *old = cl->qdisc;
622         cl->qdisc = new;
623         sch_tree_unlock(sch);
624         return 0;
625 }
626 
627 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
628 {
629         struct qfq_class *cl = (struct qfq_class *)arg;
630 
631         return cl->qdisc;
632 }
633 
634 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
635                           struct sk_buff *skb, struct tcmsg *tcm)
636 {
637         struct qfq_class *cl = (struct qfq_class *)arg;
638         struct nlattr *nest;
639 
640         tcm->tcm_parent = TC_H_ROOT;
641         tcm->tcm_handle = cl->common.classid;
642         tcm->tcm_info   = cl->qdisc->handle;
643 
644         nest = nla_nest_start(skb, TCA_OPTIONS);
645         if (nest == NULL)
646                 goto nla_put_failure;
647         if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
648             nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
649                 goto nla_put_failure;
650         return nla_nest_end(skb, nest);
651 
652 nla_put_failure:
653         nla_nest_cancel(skb, nest);
654         return -EMSGSIZE;
655 }
656 
657 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
658                                 struct gnet_dump *d)
659 {
660         struct qfq_class *cl = (struct qfq_class *)arg;
661         struct tc_qfq_stats xstats;
662 
663         memset(&xstats, 0, sizeof(xstats));
664         cl->qdisc->qstats.qlen = cl->qdisc->q.qlen;
665 
666         xstats.weight = cl->agg->class_weight;
667         xstats.lmax = cl->agg->lmax;
668 
669         if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
670             gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
671             gnet_stats_copy_queue(d, &cl->qdisc->qstats) < 0)
672                 return -1;
673 
674         return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
675 }
676 
677 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
678 {
679         struct qfq_sched *q = qdisc_priv(sch);
680         struct qfq_class *cl;
681         unsigned int i;
682 
683         if (arg->stop)
684                 return;
685 
686         for (i = 0; i < q->clhash.hashsize; i++) {
687                 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
688                         if (arg->count < arg->skip) {
689                                 arg->count++;
690                                 continue;
691                         }
692                         if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
693                                 arg->stop = 1;
694                                 return;
695                         }
696                         arg->count++;
697                 }
698         }
699 }
700 
701 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
702                                       int *qerr)
703 {
704         struct qfq_sched *q = qdisc_priv(sch);
705         struct qfq_class *cl;
706         struct tcf_result res;
707         int result;
708 
709         if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
710                 pr_debug("qfq_classify: found %d\n", skb->priority);
711                 cl = qfq_find_class(sch, skb->priority);
712                 if (cl != NULL)
713                         return cl;
714         }
715 
716         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
717         result = tc_classify(skb, q->filter_list, &res);
718         if (result >= 0) {
719 #ifdef CONFIG_NET_CLS_ACT
720                 switch (result) {
721                 case TC_ACT_QUEUED:
722                 case TC_ACT_STOLEN:
723                         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
724                 case TC_ACT_SHOT:
725                         return NULL;
726                 }
727 #endif
728                 cl = (struct qfq_class *)res.class;
729                 if (cl == NULL)
730                         cl = qfq_find_class(sch, res.classid);
731                 return cl;
732         }
733 
734         return NULL;
735 }
736 
737 /* Generic comparison function, handling wraparound. */
738 static inline int qfq_gt(u64 a, u64 b)
739 {
740         return (s64)(a - b) > 0;
741 }
742 
743 /* Round a precise timestamp to its slotted value. */
744 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
745 {
746         return ts & ~((1ULL << shift) - 1);
747 }
748 
749 /* return the pointer to the group with lowest index in the bitmap */
750 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
751                                         unsigned long bitmap)
752 {
753         int index = __ffs(bitmap);
754         return &q->groups[index];
755 }
756 /* Calculate a mask to mimic what would be ffs_from(). */
757 static inline unsigned long mask_from(unsigned long bitmap, int from)
758 {
759         return bitmap & ~((1UL << from) - 1);
760 }
761 
762 /*
763  * The state computation relies on ER=0, IR=1, EB=2, IB=3
764  * First compute eligibility comparing grp->S, q->V,
765  * then check if someone is blocking us and possibly add EB
766  */
767 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
768 {
769         /* if S > V we are not eligible */
770         unsigned int state = qfq_gt(grp->S, q->V);
771         unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
772         struct qfq_group *next;
773 
774         if (mask) {
775                 next = qfq_ffs(q, mask);
776                 if (qfq_gt(grp->F, next->F))
777                         state |= EB;
778         }
779 
780         return state;
781 }
782 
783 
784 /*
785  * In principle
786  *      q->bitmaps[dst] |= q->bitmaps[src] & mask;
787  *      q->bitmaps[src] &= ~mask;
788  * but we should make sure that src != dst
789  */
790 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
791                                    int src, int dst)
792 {
793         q->bitmaps[dst] |= q->bitmaps[src] & mask;
794         q->bitmaps[src] &= ~mask;
795 }
796 
797 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
798 {
799         unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
800         struct qfq_group *next;
801 
802         if (mask) {
803                 next = qfq_ffs(q, mask);
804                 if (!qfq_gt(next->F, old_F))
805                         return;
806         }
807 
808         mask = (1UL << index) - 1;
809         qfq_move_groups(q, mask, EB, ER);
810         qfq_move_groups(q, mask, IB, IR);
811 }
812 
813 /*
814  * perhaps
815  *
816         old_V ^= q->V;
817         old_V >>= q->min_slot_shift;
818         if (old_V) {
819                 ...
820         }
821  *
822  */
823 static void qfq_make_eligible(struct qfq_sched *q)
824 {
825         unsigned long vslot = q->V >> q->min_slot_shift;
826         unsigned long old_vslot = q->oldV >> q->min_slot_shift;
827 
828         if (vslot != old_vslot) {
829                 unsigned long mask;
830                 int last_flip_pos = fls(vslot ^ old_vslot);
831 
832                 if (last_flip_pos > 31) /* higher than the number of groups */
833                         mask = ~0UL;    /* make all groups eligible */
834                 else
835                         mask = (1UL << last_flip_pos) - 1;
836 
837                 qfq_move_groups(q, mask, IR, ER);
838                 qfq_move_groups(q, mask, IB, EB);
839         }
840 }
841 
842 /*
843  * The index of the slot in which the input aggregate agg is to be
844  * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
845  * and not a '-1' because the start time of the group may be moved
846  * backward by one slot after the aggregate has been inserted, and
847  * this would cause non-empty slots to be right-shifted by one
848  * position.
849  *
850  * QFQ+ fully satisfies this bound to the slot index if the parameters
851  * of the classes are not changed dynamically, and if QFQ+ never
852  * happens to postpone the service of agg unjustly, i.e., it never
853  * happens that the aggregate becomes backlogged and eligible, or just
854  * eligible, while an aggregate with a higher approximated finish time
855  * is being served. In particular, in this case QFQ+ guarantees that
856  * the timestamps of agg are low enough that the slot index is never
857  * higher than 2. Unfortunately, QFQ+ cannot provide the same
858  * guarantee if it happens to unjustly postpone the service of agg, or
859  * if the parameters of some class are changed.
860  *
861  * As for the first event, i.e., an out-of-order service, the
862  * upper bound to the slot index guaranteed by QFQ+ grows to
863  * 2 +
864  * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
865  * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
866  *
867  * The following function deals with this problem by backward-shifting
868  * the timestamps of agg, if needed, so as to guarantee that the slot
869  * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
870  * cause the service of other aggregates to be postponed, yet the
871  * worst-case guarantees of these aggregates are not violated.  In
872  * fact, in case of no out-of-order service, the timestamps of agg
873  * would have been even lower than they are after the backward shift,
874  * because QFQ+ would have guaranteed a maximum value equal to 2 for
875  * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
876  * service is postponed because of the backward-shift would have
877  * however waited for the service of agg before being served.
878  *
879  * The other event that may cause the slot index to be higher than 2
880  * for agg is a recent change of the parameters of some class. If the
881  * weight of a class is increased or the lmax (max_pkt_size) of the
882  * class is decreased, then a new aggregate with smaller slot size
883  * than the original parent aggregate of the class may happen to be
884  * activated. The activation of this aggregate should be properly
885  * delayed to when the service of the class has finished in the ideal
886  * system tracked by QFQ+. If the activation of the aggregate is not
887  * delayed to this reference time instant, then this aggregate may be
888  * unjustly served before other aggregates waiting for service. This
889  * may cause the above bound to the slot index to be violated for some
890  * of these unlucky aggregates.
891  *
892  * Instead of delaying the activation of the new aggregate, which is
893  * quite complex, the above-discussed capping of the slot index is
894  * used to handle also the consequences of a change of the parameters
895  * of a class.
896  */
897 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
898                             u64 roundedS)
899 {
900         u64 slot = (roundedS - grp->S) >> grp->slot_shift;
901         unsigned int i; /* slot index in the bucket list */
902 
903         if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
904                 u64 deltaS = roundedS - grp->S -
905                         ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
906                 agg->S -= deltaS;
907                 agg->F -= deltaS;
908                 slot = QFQ_MAX_SLOTS - 2;
909         }
910 
911         i = (grp->front + slot) % QFQ_MAX_SLOTS;
912 
913         hlist_add_head(&agg->next, &grp->slots[i]);
914         __set_bit(slot, &grp->full_slots);
915 }
916 
917 /* Maybe introduce hlist_first_entry?? */
918 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
919 {
920         return hlist_entry(grp->slots[grp->front].first,
921                            struct qfq_aggregate, next);
922 }
923 
924 /*
925  * remove the entry from the slot
926  */
927 static void qfq_front_slot_remove(struct qfq_group *grp)
928 {
929         struct qfq_aggregate *agg = qfq_slot_head(grp);
930 
931         BUG_ON(!agg);
932         hlist_del(&agg->next);
933         if (hlist_empty(&grp->slots[grp->front]))
934                 __clear_bit(0, &grp->full_slots);
935 }
936 
937 /*
938  * Returns the first aggregate in the first non-empty bucket of the
939  * group. As a side effect, adjusts the bucket list so the first
940  * non-empty bucket is at position 0 in full_slots.
941  */
942 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
943 {
944         unsigned int i;
945 
946         pr_debug("qfq slot_scan: grp %u full %#lx\n",
947                  grp->index, grp->full_slots);
948 
949         if (grp->full_slots == 0)
950                 return NULL;
951 
952         i = __ffs(grp->full_slots);  /* zero based */
953         if (i > 0) {
954                 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
955                 grp->full_slots >>= i;
956         }
957 
958         return qfq_slot_head(grp);
959 }
960 
961 /*
962  * adjust the bucket list. When the start time of a group decreases,
963  * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
964  * move the objects. The mask of occupied slots must be shifted
965  * because we use ffs() to find the first non-empty slot.
966  * This covers decreases in the group's start time, but what about
967  * increases of the start time ?
968  * Here too we should make sure that i is less than 32
969  */
970 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
971 {
972         unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
973 
974         grp->full_slots <<= i;
975         grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
976 }
977 
978 static void qfq_update_eligible(struct qfq_sched *q)
979 {
980         struct qfq_group *grp;
981         unsigned long ineligible;
982 
983         ineligible = q->bitmaps[IR] | q->bitmaps[IB];
984         if (ineligible) {
985                 if (!q->bitmaps[ER]) {
986                         grp = qfq_ffs(q, ineligible);
987                         if (qfq_gt(grp->S, q->V))
988                                 q->V = grp->S;
989                 }
990                 qfq_make_eligible(q);
991         }
992 }
993 
994 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
995 static void agg_dequeue(struct qfq_aggregate *agg,
996                         struct qfq_class *cl, unsigned int len)
997 {
998         qdisc_dequeue_peeked(cl->qdisc);
999 
1000         cl->deficit -= (int) len;
1001 
1002         if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
1003                 list_del(&cl->alist);
1004         else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
1005                 cl->deficit += agg->lmax;
1006                 list_move_tail(&cl->alist, &agg->active);
1007         }
1008 }
1009 
1010 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
1011                                            struct qfq_class **cl,
1012                                            unsigned int *len)
1013 {
1014         struct sk_buff *skb;
1015 
1016         *cl = list_first_entry(&agg->active, struct qfq_class, alist);
1017         skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
1018         if (skb == NULL)
1019                 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
1020         else
1021                 *len = qdisc_pkt_len(skb);
1022 
1023         return skb;
1024 }
1025 
1026 /* Update F according to the actual service received by the aggregate. */
1027 static inline void charge_actual_service(struct qfq_aggregate *agg)
1028 {
1029         /* Compute the service received by the aggregate, taking into
1030          * account that, after decreasing the number of classes in
1031          * agg, it may happen that
1032          * agg->initial_budget - agg->budget > agg->bugdetmax
1033          */
1034         u32 service_received = min(agg->budgetmax,
1035                                    agg->initial_budget - agg->budget);
1036 
1037         agg->F = agg->S + (u64)service_received * agg->inv_w;
1038 }
1039 
1040 /* Assign a reasonable start time for a new aggregate in group i.
1041  * Admissible values for \hat(F) are multiples of \sigma_i
1042  * no greater than V+\sigma_i . Larger values mean that
1043  * we had a wraparound so we consider the timestamp to be stale.
1044  *
1045  * If F is not stale and F >= V then we set S = F.
1046  * Otherwise we should assign S = V, but this may violate
1047  * the ordering in EB (see [2]). So, if we have groups in ER,
1048  * set S to the F_j of the first group j which would be blocking us.
1049  * We are guaranteed not to move S backward because
1050  * otherwise our group i would still be blocked.
1051  */
1052 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1053 {
1054         unsigned long mask;
1055         u64 limit, roundedF;
1056         int slot_shift = agg->grp->slot_shift;
1057 
1058         roundedF = qfq_round_down(agg->F, slot_shift);
1059         limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1060 
1061         if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1062                 /* timestamp was stale */
1063                 mask = mask_from(q->bitmaps[ER], agg->grp->index);
1064                 if (mask) {
1065                         struct qfq_group *next = qfq_ffs(q, mask);
1066                         if (qfq_gt(roundedF, next->F)) {
1067                                 if (qfq_gt(limit, next->F))
1068                                         agg->S = next->F;
1069                                 else /* preserve timestamp correctness */
1070                                         agg->S = limit;
1071                                 return;
1072                         }
1073                 }
1074                 agg->S = q->V;
1075         } else  /* timestamp is not stale */
1076                 agg->S = agg->F;
1077 }
1078 
1079 /* Update the timestamps of agg before scheduling/rescheduling it for
1080  * service.  In particular, assign to agg->F its maximum possible
1081  * value, i.e., the virtual finish time with which the aggregate
1082  * should be labeled if it used all its budget once in service.
1083  */
1084 static inline void
1085 qfq_update_agg_ts(struct qfq_sched *q,
1086                     struct qfq_aggregate *agg, enum update_reason reason)
1087 {
1088         if (reason != requeue)
1089                 qfq_update_start(q, agg);
1090         else /* just charge agg for the service received */
1091                 agg->S = agg->F;
1092 
1093         agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1094 }
1095 
1096 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1097 
1098 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1099 {
1100         struct qfq_sched *q = qdisc_priv(sch);
1101         struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1102         struct qfq_class *cl;
1103         struct sk_buff *skb = NULL;
1104         /* next-packet len, 0 means no more active classes in in-service agg */
1105         unsigned int len = 0;
1106 
1107         if (in_serv_agg == NULL)
1108                 return NULL;
1109 
1110         if (!list_empty(&in_serv_agg->active))
1111                 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1112 
1113         /*
1114          * If there are no active classes in the in-service aggregate,
1115          * or if the aggregate has not enough budget to serve its next
1116          * class, then choose the next aggregate to serve.
1117          */
1118         if (len == 0 || in_serv_agg->budget < len) {
1119                 charge_actual_service(in_serv_agg);
1120 
1121                 /* recharge the budget of the aggregate */
1122                 in_serv_agg->initial_budget = in_serv_agg->budget =
1123                         in_serv_agg->budgetmax;
1124 
1125                 if (!list_empty(&in_serv_agg->active)) {
1126                         /*
1127                          * Still active: reschedule for
1128                          * service. Possible optimization: if no other
1129                          * aggregate is active, then there is no point
1130                          * in rescheduling this aggregate, and we can
1131                          * just keep it as the in-service one. This
1132                          * should be however a corner case, and to
1133                          * handle it, we would need to maintain an
1134                          * extra num_active_aggs field.
1135                         */
1136                         qfq_update_agg_ts(q, in_serv_agg, requeue);
1137                         qfq_schedule_agg(q, in_serv_agg);
1138                 } else if (sch->q.qlen == 0) { /* no aggregate to serve */
1139                         q->in_serv_agg = NULL;
1140                         return NULL;
1141                 }
1142 
1143                 /*
1144                  * If we get here, there are other aggregates queued:
1145                  * choose the new aggregate to serve.
1146                  */
1147                 in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1148                 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1149         }
1150         if (!skb)
1151                 return NULL;
1152 
1153         sch->q.qlen--;
1154         qdisc_bstats_update(sch, skb);
1155 
1156         agg_dequeue(in_serv_agg, cl, len);
1157         /* If lmax is lowered, through qfq_change_class, for a class
1158          * owning pending packets with larger size than the new value
1159          * of lmax, then the following condition may hold.
1160          */
1161         if (unlikely(in_serv_agg->budget < len))
1162                 in_serv_agg->budget = 0;
1163         else
1164                 in_serv_agg->budget -= len;
1165 
1166         q->V += (u64)len * q->iwsum;
1167         pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1168                  len, (unsigned long long) in_serv_agg->F,
1169                  (unsigned long long) q->V);
1170 
1171         return skb;
1172 }
1173 
1174 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1175 {
1176         struct qfq_group *grp;
1177         struct qfq_aggregate *agg, *new_front_agg;
1178         u64 old_F;
1179 
1180         qfq_update_eligible(q);
1181         q->oldV = q->V;
1182 
1183         if (!q->bitmaps[ER])
1184                 return NULL;
1185 
1186         grp = qfq_ffs(q, q->bitmaps[ER]);
1187         old_F = grp->F;
1188 
1189         agg = qfq_slot_head(grp);
1190 
1191         /* agg starts to be served, remove it from schedule */
1192         qfq_front_slot_remove(grp);
1193 
1194         new_front_agg = qfq_slot_scan(grp);
1195 
1196         if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1197                 __clear_bit(grp->index, &q->bitmaps[ER]);
1198         else {
1199                 u64 roundedS = qfq_round_down(new_front_agg->S,
1200                                               grp->slot_shift);
1201                 unsigned int s;
1202 
1203                 if (grp->S == roundedS)
1204                         return agg;
1205                 grp->S = roundedS;
1206                 grp->F = roundedS + (2ULL << grp->slot_shift);
1207                 __clear_bit(grp->index, &q->bitmaps[ER]);
1208                 s = qfq_calc_state(q, grp);
1209                 __set_bit(grp->index, &q->bitmaps[s]);
1210         }
1211 
1212         qfq_unblock_groups(q, grp->index, old_F);
1213 
1214         return agg;
1215 }
1216 
1217 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1218 {
1219         struct qfq_sched *q = qdisc_priv(sch);
1220         struct qfq_class *cl;
1221         struct qfq_aggregate *agg;
1222         int err = 0;
1223 
1224         cl = qfq_classify(skb, sch, &err);
1225         if (cl == NULL) {
1226                 if (err & __NET_XMIT_BYPASS)
1227                         sch->qstats.drops++;
1228                 kfree_skb(skb);
1229                 return err;
1230         }
1231         pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1232 
1233         if (unlikely(cl->agg->lmax < qdisc_pkt_len(skb))) {
1234                 pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1235                          cl->agg->lmax, qdisc_pkt_len(skb), cl->common.classid);
1236                 err = qfq_change_agg(sch, cl, cl->agg->class_weight,
1237                                      qdisc_pkt_len(skb));
1238                 if (err)
1239                         return err;
1240         }
1241 
1242         err = qdisc_enqueue(skb, cl->qdisc);
1243         if (unlikely(err != NET_XMIT_SUCCESS)) {
1244                 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1245                 if (net_xmit_drop_count(err)) {
1246                         cl->qstats.drops++;
1247                         sch->qstats.drops++;
1248                 }
1249                 return err;
1250         }
1251 
1252         bstats_update(&cl->bstats, skb);
1253         ++sch->q.qlen;
1254 
1255         agg = cl->agg;
1256         /* if the queue was not empty, then done here */
1257         if (cl->qdisc->q.qlen != 1) {
1258                 if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1259                     list_first_entry(&agg->active, struct qfq_class, alist)
1260                     == cl && cl->deficit < qdisc_pkt_len(skb))
1261                         list_move_tail(&cl->alist, &agg->active);
1262 
1263                 return err;
1264         }
1265 
1266         /* schedule class for service within the aggregate */
1267         cl->deficit = agg->lmax;
1268         list_add_tail(&cl->alist, &agg->active);
1269 
1270         if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1271             q->in_serv_agg == agg)
1272                 return err; /* non-empty or in service, nothing else to do */
1273 
1274         qfq_activate_agg(q, agg, enqueue);
1275 
1276         return err;
1277 }
1278 
1279 /*
1280  * Schedule aggregate according to its timestamps.
1281  */
1282 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1283 {
1284         struct qfq_group *grp = agg->grp;
1285         u64 roundedS;
1286         int s;
1287 
1288         roundedS = qfq_round_down(agg->S, grp->slot_shift);
1289 
1290         /*
1291          * Insert agg in the correct bucket.
1292          * If agg->S >= grp->S we don't need to adjust the
1293          * bucket list and simply go to the insertion phase.
1294          * Otherwise grp->S is decreasing, we must make room
1295          * in the bucket list, and also recompute the group state.
1296          * Finally, if there were no flows in this group and nobody
1297          * was in ER make sure to adjust V.
1298          */
1299         if (grp->full_slots) {
1300                 if (!qfq_gt(grp->S, agg->S))
1301                         goto skip_update;
1302 
1303                 /* create a slot for this agg->S */
1304                 qfq_slot_rotate(grp, roundedS);
1305                 /* group was surely ineligible, remove */
1306                 __clear_bit(grp->index, &q->bitmaps[IR]);
1307                 __clear_bit(grp->index, &q->bitmaps[IB]);
1308         } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1309                    q->in_serv_agg == NULL)
1310                 q->V = roundedS;
1311 
1312         grp->S = roundedS;
1313         grp->F = roundedS + (2ULL << grp->slot_shift);
1314         s = qfq_calc_state(q, grp);
1315         __set_bit(grp->index, &q->bitmaps[s]);
1316 
1317         pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1318                  s, q->bitmaps[s],
1319                  (unsigned long long) agg->S,
1320                  (unsigned long long) agg->F,
1321                  (unsigned long long) q->V);
1322 
1323 skip_update:
1324         qfq_slot_insert(grp, agg, roundedS);
1325 }
1326 
1327 
1328 /* Update agg ts and schedule agg for service */
1329 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1330                              enum update_reason reason)
1331 {
1332         agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1333 
1334         qfq_update_agg_ts(q, agg, reason);
1335         if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1336                 q->in_serv_agg = agg; /* start serving this aggregate */
1337                  /* update V: to be in service, agg must be eligible */
1338                 q->oldV = q->V = agg->S;
1339         } else if (agg != q->in_serv_agg)
1340                 qfq_schedule_agg(q, agg);
1341 }
1342 
1343 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1344                             struct qfq_aggregate *agg)
1345 {
1346         unsigned int i, offset;
1347         u64 roundedS;
1348 
1349         roundedS = qfq_round_down(agg->S, grp->slot_shift);
1350         offset = (roundedS - grp->S) >> grp->slot_shift;
1351 
1352         i = (grp->front + offset) % QFQ_MAX_SLOTS;
1353 
1354         hlist_del(&agg->next);
1355         if (hlist_empty(&grp->slots[i]))
1356                 __clear_bit(offset, &grp->full_slots);
1357 }
1358 
1359 /*
1360  * Called to forcibly deschedule an aggregate.  If the aggregate is
1361  * not in the front bucket, or if the latter has other aggregates in
1362  * the front bucket, we can simply remove the aggregate with no other
1363  * side effects.
1364  * Otherwise we must propagate the event up.
1365  */
1366 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1367 {
1368         struct qfq_group *grp = agg->grp;
1369         unsigned long mask;
1370         u64 roundedS;
1371         int s;
1372 
1373         if (agg == q->in_serv_agg) {
1374                 charge_actual_service(agg);
1375                 q->in_serv_agg = qfq_choose_next_agg(q);
1376                 return;
1377         }
1378 
1379         agg->F = agg->S;
1380         qfq_slot_remove(q, grp, agg);
1381 
1382         if (!grp->full_slots) {
1383                 __clear_bit(grp->index, &q->bitmaps[IR]);
1384                 __clear_bit(grp->index, &q->bitmaps[EB]);
1385                 __clear_bit(grp->index, &q->bitmaps[IB]);
1386 
1387                 if (test_bit(grp->index, &q->bitmaps[ER]) &&
1388                     !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1389                         mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1390                         if (mask)
1391                                 mask = ~((1UL << __fls(mask)) - 1);
1392                         else
1393                                 mask = ~0UL;
1394                         qfq_move_groups(q, mask, EB, ER);
1395                         qfq_move_groups(q, mask, IB, IR);
1396                 }
1397                 __clear_bit(grp->index, &q->bitmaps[ER]);
1398         } else if (hlist_empty(&grp->slots[grp->front])) {
1399                 agg = qfq_slot_scan(grp);
1400                 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1401                 if (grp->S != roundedS) {
1402                         __clear_bit(grp->index, &q->bitmaps[ER]);
1403                         __clear_bit(grp->index, &q->bitmaps[IR]);
1404                         __clear_bit(grp->index, &q->bitmaps[EB]);
1405                         __clear_bit(grp->index, &q->bitmaps[IB]);
1406                         grp->S = roundedS;
1407                         grp->F = roundedS + (2ULL << grp->slot_shift);
1408                         s = qfq_calc_state(q, grp);
1409                         __set_bit(grp->index, &q->bitmaps[s]);
1410                 }
1411         }
1412 }
1413 
1414 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1415 {
1416         struct qfq_sched *q = qdisc_priv(sch);
1417         struct qfq_class *cl = (struct qfq_class *)arg;
1418 
1419         if (cl->qdisc->q.qlen == 0)
1420                 qfq_deactivate_class(q, cl);
1421 }
1422 
1423 static unsigned int qfq_drop_from_slot(struct qfq_sched *q,
1424                                        struct hlist_head *slot)
1425 {
1426         struct qfq_aggregate *agg;
1427         struct qfq_class *cl;
1428         unsigned int len;
1429 
1430         hlist_for_each_entry(agg, slot, next) {
1431                 list_for_each_entry(cl, &agg->active, alist) {
1432 
1433                         if (!cl->qdisc->ops->drop)
1434                                 continue;
1435 
1436                         len = cl->qdisc->ops->drop(cl->qdisc);
1437                         if (len > 0) {
1438                                 if (cl->qdisc->q.qlen == 0)
1439                                         qfq_deactivate_class(q, cl);
1440 
1441                                 return len;
1442                         }
1443                 }
1444         }
1445         return 0;
1446 }
1447 
1448 static unsigned int qfq_drop(struct Qdisc *sch)
1449 {
1450         struct qfq_sched *q = qdisc_priv(sch);
1451         struct qfq_group *grp;
1452         unsigned int i, j, len;
1453 
1454         for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1455                 grp = &q->groups[i];
1456                 for (j = 0; j < QFQ_MAX_SLOTS; j++) {
1457                         len = qfq_drop_from_slot(q, &grp->slots[j]);
1458                         if (len > 0) {
1459                                 sch->q.qlen--;
1460                                 return len;
1461                         }
1462                 }
1463 
1464         }
1465 
1466         return 0;
1467 }
1468 
1469 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1470 {
1471         struct qfq_sched *q = qdisc_priv(sch);
1472         struct qfq_group *grp;
1473         int i, j, err;
1474         u32 max_cl_shift, maxbudg_shift, max_classes;
1475 
1476         err = qdisc_class_hash_init(&q->clhash);
1477         if (err < 0)
1478                 return err;
1479 
1480         if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES)
1481                 max_classes = QFQ_MAX_AGG_CLASSES;
1482         else
1483                 max_classes = qdisc_dev(sch)->tx_queue_len + 1;
1484         /* max_cl_shift = floor(log_2(max_classes)) */
1485         max_cl_shift = __fls(max_classes);
1486         q->max_agg_classes = 1<<max_cl_shift;
1487 
1488         /* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1489         maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1490         q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1491 
1492         for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1493                 grp = &q->groups[i];
1494                 grp->index = i;
1495                 grp->slot_shift = q->min_slot_shift + i;
1496                 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1497                         INIT_HLIST_HEAD(&grp->slots[j]);
1498         }
1499 
1500         INIT_HLIST_HEAD(&q->nonfull_aggs);
1501 
1502         return 0;
1503 }
1504 
1505 static void qfq_reset_qdisc(struct Qdisc *sch)
1506 {
1507         struct qfq_sched *q = qdisc_priv(sch);
1508         struct qfq_class *cl;
1509         unsigned int i;
1510 
1511         for (i = 0; i < q->clhash.hashsize; i++) {
1512                 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1513                         if (cl->qdisc->q.qlen > 0)
1514                                 qfq_deactivate_class(q, cl);
1515 
1516                         qdisc_reset(cl->qdisc);
1517                 }
1518         }
1519         sch->q.qlen = 0;
1520 }
1521 
1522 static void qfq_destroy_qdisc(struct Qdisc *sch)
1523 {
1524         struct qfq_sched *q = qdisc_priv(sch);
1525         struct qfq_class *cl;
1526         struct hlist_node *next;
1527         unsigned int i;
1528 
1529         tcf_destroy_chain(&q->filter_list);
1530 
1531         for (i = 0; i < q->clhash.hashsize; i++) {
1532                 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1533                                           common.hnode) {
1534                         qfq_destroy_class(sch, cl);
1535                 }
1536         }
1537         qdisc_class_hash_destroy(&q->clhash);
1538 }
1539 
1540 static const struct Qdisc_class_ops qfq_class_ops = {
1541         .change         = qfq_change_class,
1542         .delete         = qfq_delete_class,
1543         .get            = qfq_get_class,
1544         .put            = qfq_put_class,
1545         .tcf_chain      = qfq_tcf_chain,
1546         .bind_tcf       = qfq_bind_tcf,
1547         .unbind_tcf     = qfq_unbind_tcf,
1548         .graft          = qfq_graft_class,
1549         .leaf           = qfq_class_leaf,
1550         .qlen_notify    = qfq_qlen_notify,
1551         .dump           = qfq_dump_class,
1552         .dump_stats     = qfq_dump_class_stats,
1553         .walk           = qfq_walk,
1554 };
1555 
1556 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1557         .cl_ops         = &qfq_class_ops,
1558         .id             = "qfq",
1559         .priv_size      = sizeof(struct qfq_sched),
1560         .enqueue        = qfq_enqueue,
1561         .dequeue        = qfq_dequeue,
1562         .peek           = qdisc_peek_dequeued,
1563         .drop           = qfq_drop,
1564         .init           = qfq_init_qdisc,
1565         .reset          = qfq_reset_qdisc,
1566         .destroy        = qfq_destroy_qdisc,
1567         .owner          = THIS_MODULE,
1568 };
1569 
1570 static int __init qfq_init(void)
1571 {
1572         return register_qdisc(&qfq_qdisc_ops);
1573 }
1574 
1575 static void __exit qfq_exit(void)
1576 {
1577         unregister_qdisc(&qfq_qdisc_ops);
1578 }
1579 
1580 module_init(qfq_init);
1581 module_exit(qfq_exit);
1582 MODULE_LICENSE("GPL");
1583 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp