Merge git://git.kernel.org/pub/scm/linux/kernel/git/pablo/nf-next
[firefly-linux-kernel-4.4.55.git] / net / sched / sch_fq_codel.c
1 /*
2  * Fair Queue CoDel discipline
3  *
4  *      This program is free software; you can redistribute it and/or
5  *      modify it under the terms of the GNU General Public License
6  *      as published by the Free Software Foundation; either version
7  *      2 of the License, or (at your option) any later version.
8  *
9  *  Copyright (C) 2012 Eric Dumazet <edumazet@google.com>
10  */
11
12 #include <linux/module.h>
13 #include <linux/types.h>
14 #include <linux/kernel.h>
15 #include <linux/jiffies.h>
16 #include <linux/string.h>
17 #include <linux/in.h>
18 #include <linux/errno.h>
19 #include <linux/init.h>
20 #include <linux/skbuff.h>
21 #include <linux/jhash.h>
22 #include <linux/slab.h>
23 #include <linux/vmalloc.h>
24 #include <net/netlink.h>
25 #include <net/pkt_sched.h>
26 #include <net/flow_keys.h>
27 #include <net/codel.h>
28
29 /*      Fair Queue CoDel.
30  *
31  * Principles :
32  * Packets are classified (internal classifier or external) on flows.
33  * This is a Stochastic model (as we use a hash, several flows
34  *                             might be hashed on same slot)
35  * Each flow has a CoDel managed queue.
36  * Flows are linked onto two (Round Robin) lists,
37  * so that new flows have priority on old ones.
38  *
39  * For a given flow, packets are not reordered (CoDel uses a FIFO)
40  * head drops only.
41  * ECN capability is on by default.
42  * Low memory footprint (64 bytes per flow)
43  */
44
45 struct fq_codel_flow {
46         struct sk_buff    *head;
47         struct sk_buff    *tail;
48         struct list_head  flowchain;
49         int               deficit;
50         u32               dropped; /* number of drops (or ECN marks) on this flow */
51         struct codel_vars cvars;
52 }; /* please try to keep this structure <= 64 bytes */
53
54 struct fq_codel_sched_data {
55         struct tcf_proto *filter_list;  /* optional external classifier */
56         struct fq_codel_flow *flows;    /* Flows table [flows_cnt] */
57         u32             *backlogs;      /* backlog table [flows_cnt] */
58         u32             flows_cnt;      /* number of flows */
59         u32             perturbation;   /* hash perturbation */
60         u32             quantum;        /* psched_mtu(qdisc_dev(sch)); */
61         struct codel_params cparams;
62         struct codel_stats cstats;
63         u32             drop_overlimit;
64         u32             new_flow_count;
65
66         struct list_head new_flows;     /* list of new flows */
67         struct list_head old_flows;     /* list of old flows */
68 };
69
70 static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
71                                   const struct sk_buff *skb)
72 {
73         struct flow_keys keys;
74         unsigned int hash;
75
76         skb_flow_dissect(skb, &keys);
77         hash = jhash_3words((__force u32)keys.dst,
78                             (__force u32)keys.src ^ keys.ip_proto,
79                             (__force u32)keys.ports, q->perturbation);
80
81         return reciprocal_scale(hash, q->flows_cnt);
82 }
83
84 static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
85                                       int *qerr)
86 {
87         struct fq_codel_sched_data *q = qdisc_priv(sch);
88         struct tcf_result res;
89         int result;
90
91         if (TC_H_MAJ(skb->priority) == sch->handle &&
92             TC_H_MIN(skb->priority) > 0 &&
93             TC_H_MIN(skb->priority) <= q->flows_cnt)
94                 return TC_H_MIN(skb->priority);
95
96         if (!q->filter_list)
97                 return fq_codel_hash(q, skb) + 1;
98
99         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
100         result = tc_classify(skb, q->filter_list, &res);
101         if (result >= 0) {
102 #ifdef CONFIG_NET_CLS_ACT
103                 switch (result) {
104                 case TC_ACT_STOLEN:
105                 case TC_ACT_QUEUED:
106                         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
107                 case TC_ACT_SHOT:
108                         return 0;
109                 }
110 #endif
111                 if (TC_H_MIN(res.classid) <= q->flows_cnt)
112                         return TC_H_MIN(res.classid);
113         }
114         return 0;
115 }
116
117 /* helper functions : might be changed when/if skb use a standard list_head */
118
119 /* remove one skb from head of slot queue */
120 static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
121 {
122         struct sk_buff *skb = flow->head;
123
124         flow->head = skb->next;
125         skb->next = NULL;
126         return skb;
127 }
128
129 /* add skb to flow queue (tail add) */
130 static inline void flow_queue_add(struct fq_codel_flow *flow,
131                                   struct sk_buff *skb)
132 {
133         if (flow->head == NULL)
134                 flow->head = skb;
135         else
136                 flow->tail->next = skb;
137         flow->tail = skb;
138         skb->next = NULL;
139 }
140
141 static unsigned int fq_codel_drop(struct Qdisc *sch)
142 {
143         struct fq_codel_sched_data *q = qdisc_priv(sch);
144         struct sk_buff *skb;
145         unsigned int maxbacklog = 0, idx = 0, i, len;
146         struct fq_codel_flow *flow;
147
148         /* Queue is full! Find the fat flow and drop packet from it.
149          * This might sound expensive, but with 1024 flows, we scan
150          * 4KB of memory, and we dont need to handle a complex tree
151          * in fast path (packet queue/enqueue) with many cache misses.
152          */
153         for (i = 0; i < q->flows_cnt; i++) {
154                 if (q->backlogs[i] > maxbacklog) {
155                         maxbacklog = q->backlogs[i];
156                         idx = i;
157                 }
158         }
159         flow = &q->flows[idx];
160         skb = dequeue_head(flow);
161         len = qdisc_pkt_len(skb);
162         q->backlogs[idx] -= len;
163         kfree_skb(skb);
164         sch->q.qlen--;
165         sch->qstats.drops++;
166         sch->qstats.backlog -= len;
167         flow->dropped++;
168         return idx;
169 }
170
171 static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch)
172 {
173         struct fq_codel_sched_data *q = qdisc_priv(sch);
174         unsigned int idx;
175         struct fq_codel_flow *flow;
176         int uninitialized_var(ret);
177
178         idx = fq_codel_classify(skb, sch, &ret);
179         if (idx == 0) {
180                 if (ret & __NET_XMIT_BYPASS)
181                         sch->qstats.drops++;
182                 kfree_skb(skb);
183                 return ret;
184         }
185         idx--;
186
187         codel_set_enqueue_time(skb);
188         flow = &q->flows[idx];
189         flow_queue_add(flow, skb);
190         q->backlogs[idx] += qdisc_pkt_len(skb);
191         sch->qstats.backlog += qdisc_pkt_len(skb);
192
193         if (list_empty(&flow->flowchain)) {
194                 list_add_tail(&flow->flowchain, &q->new_flows);
195                 q->new_flow_count++;
196                 flow->deficit = q->quantum;
197                 flow->dropped = 0;
198         }
199         if (++sch->q.qlen <= sch->limit)
200                 return NET_XMIT_SUCCESS;
201
202         q->drop_overlimit++;
203         /* Return Congestion Notification only if we dropped a packet
204          * from this flow.
205          */
206         if (fq_codel_drop(sch) == idx)
207                 return NET_XMIT_CN;
208
209         /* As we dropped a packet, better let upper stack know this */
210         qdisc_tree_decrease_qlen(sch, 1);
211         return NET_XMIT_SUCCESS;
212 }
213
214 /* This is the specific function called from codel_dequeue()
215  * to dequeue a packet from queue. Note: backlog is handled in
216  * codel, we dont need to reduce it here.
217  */
218 static struct sk_buff *dequeue(struct codel_vars *vars, struct Qdisc *sch)
219 {
220         struct fq_codel_sched_data *q = qdisc_priv(sch);
221         struct fq_codel_flow *flow;
222         struct sk_buff *skb = NULL;
223
224         flow = container_of(vars, struct fq_codel_flow, cvars);
225         if (flow->head) {
226                 skb = dequeue_head(flow);
227                 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
228                 sch->q.qlen--;
229         }
230         return skb;
231 }
232
233 static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
234 {
235         struct fq_codel_sched_data *q = qdisc_priv(sch);
236         struct sk_buff *skb;
237         struct fq_codel_flow *flow;
238         struct list_head *head;
239         u32 prev_drop_count, prev_ecn_mark;
240
241 begin:
242         head = &q->new_flows;
243         if (list_empty(head)) {
244                 head = &q->old_flows;
245                 if (list_empty(head))
246                         return NULL;
247         }
248         flow = list_first_entry(head, struct fq_codel_flow, flowchain);
249
250         if (flow->deficit <= 0) {
251                 flow->deficit += q->quantum;
252                 list_move_tail(&flow->flowchain, &q->old_flows);
253                 goto begin;
254         }
255
256         prev_drop_count = q->cstats.drop_count;
257         prev_ecn_mark = q->cstats.ecn_mark;
258
259         skb = codel_dequeue(sch, &q->cparams, &flow->cvars, &q->cstats,
260                             dequeue);
261
262         flow->dropped += q->cstats.drop_count - prev_drop_count;
263         flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
264
265         if (!skb) {
266                 /* force a pass through old_flows to prevent starvation */
267                 if ((head == &q->new_flows) && !list_empty(&q->old_flows))
268                         list_move_tail(&flow->flowchain, &q->old_flows);
269                 else
270                         list_del_init(&flow->flowchain);
271                 goto begin;
272         }
273         qdisc_bstats_update(sch, skb);
274         flow->deficit -= qdisc_pkt_len(skb);
275         /* We cant call qdisc_tree_decrease_qlen() if our qlen is 0,
276          * or HTB crashes. Defer it for next round.
277          */
278         if (q->cstats.drop_count && sch->q.qlen) {
279                 qdisc_tree_decrease_qlen(sch, q->cstats.drop_count);
280                 q->cstats.drop_count = 0;
281         }
282         return skb;
283 }
284
285 static void fq_codel_reset(struct Qdisc *sch)
286 {
287         struct sk_buff *skb;
288
289         while ((skb = fq_codel_dequeue(sch)) != NULL)
290                 kfree_skb(skb);
291 }
292
293 static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
294         [TCA_FQ_CODEL_TARGET]   = { .type = NLA_U32 },
295         [TCA_FQ_CODEL_LIMIT]    = { .type = NLA_U32 },
296         [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
297         [TCA_FQ_CODEL_ECN]      = { .type = NLA_U32 },
298         [TCA_FQ_CODEL_FLOWS]    = { .type = NLA_U32 },
299         [TCA_FQ_CODEL_QUANTUM]  = { .type = NLA_U32 },
300 };
301
302 static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt)
303 {
304         struct fq_codel_sched_data *q = qdisc_priv(sch);
305         struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
306         int err;
307
308         if (!opt)
309                 return -EINVAL;
310
311         err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy);
312         if (err < 0)
313                 return err;
314         if (tb[TCA_FQ_CODEL_FLOWS]) {
315                 if (q->flows)
316                         return -EINVAL;
317                 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
318                 if (!q->flows_cnt ||
319                     q->flows_cnt > 65536)
320                         return -EINVAL;
321         }
322         sch_tree_lock(sch);
323
324         if (tb[TCA_FQ_CODEL_TARGET]) {
325                 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
326
327                 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
328         }
329
330         if (tb[TCA_FQ_CODEL_INTERVAL]) {
331                 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
332
333                 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
334         }
335
336         if (tb[TCA_FQ_CODEL_LIMIT])
337                 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
338
339         if (tb[TCA_FQ_CODEL_ECN])
340                 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
341
342         if (tb[TCA_FQ_CODEL_QUANTUM])
343                 q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
344
345         while (sch->q.qlen > sch->limit) {
346                 struct sk_buff *skb = fq_codel_dequeue(sch);
347
348                 kfree_skb(skb);
349                 q->cstats.drop_count++;
350         }
351         qdisc_tree_decrease_qlen(sch, q->cstats.drop_count);
352         q->cstats.drop_count = 0;
353
354         sch_tree_unlock(sch);
355         return 0;
356 }
357
358 static void *fq_codel_zalloc(size_t sz)
359 {
360         void *ptr = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
361
362         if (!ptr)
363                 ptr = vzalloc(sz);
364         return ptr;
365 }
366
367 static void fq_codel_free(void *addr)
368 {
369         kvfree(addr);
370 }
371
372 static void fq_codel_destroy(struct Qdisc *sch)
373 {
374         struct fq_codel_sched_data *q = qdisc_priv(sch);
375
376         tcf_destroy_chain(&q->filter_list);
377         fq_codel_free(q->backlogs);
378         fq_codel_free(q->flows);
379 }
380
381 static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt)
382 {
383         struct fq_codel_sched_data *q = qdisc_priv(sch);
384         int i;
385
386         sch->limit = 10*1024;
387         q->flows_cnt = 1024;
388         q->quantum = psched_mtu(qdisc_dev(sch));
389         q->perturbation = prandom_u32();
390         INIT_LIST_HEAD(&q->new_flows);
391         INIT_LIST_HEAD(&q->old_flows);
392         codel_params_init(&q->cparams);
393         codel_stats_init(&q->cstats);
394         q->cparams.ecn = true;
395
396         if (opt) {
397                 int err = fq_codel_change(sch, opt);
398                 if (err)
399                         return err;
400         }
401
402         if (!q->flows) {
403                 q->flows = fq_codel_zalloc(q->flows_cnt *
404                                            sizeof(struct fq_codel_flow));
405                 if (!q->flows)
406                         return -ENOMEM;
407                 q->backlogs = fq_codel_zalloc(q->flows_cnt * sizeof(u32));
408                 if (!q->backlogs) {
409                         fq_codel_free(q->flows);
410                         return -ENOMEM;
411                 }
412                 for (i = 0; i < q->flows_cnt; i++) {
413                         struct fq_codel_flow *flow = q->flows + i;
414
415                         INIT_LIST_HEAD(&flow->flowchain);
416                         codel_vars_init(&flow->cvars);
417                 }
418         }
419         if (sch->limit >= 1)
420                 sch->flags |= TCQ_F_CAN_BYPASS;
421         else
422                 sch->flags &= ~TCQ_F_CAN_BYPASS;
423         return 0;
424 }
425
426 static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
427 {
428         struct fq_codel_sched_data *q = qdisc_priv(sch);
429         struct nlattr *opts;
430
431         opts = nla_nest_start(skb, TCA_OPTIONS);
432         if (opts == NULL)
433                 goto nla_put_failure;
434
435         if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
436                         codel_time_to_us(q->cparams.target)) ||
437             nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
438                         sch->limit) ||
439             nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
440                         codel_time_to_us(q->cparams.interval)) ||
441             nla_put_u32(skb, TCA_FQ_CODEL_ECN,
442                         q->cparams.ecn) ||
443             nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
444                         q->quantum) ||
445             nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
446                         q->flows_cnt))
447                 goto nla_put_failure;
448
449         return nla_nest_end(skb, opts);
450
451 nla_put_failure:
452         return -1;
453 }
454
455 static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
456 {
457         struct fq_codel_sched_data *q = qdisc_priv(sch);
458         struct tc_fq_codel_xstats st = {
459                 .type                           = TCA_FQ_CODEL_XSTATS_QDISC,
460         };
461         struct list_head *pos;
462
463         st.qdisc_stats.maxpacket = q->cstats.maxpacket;
464         st.qdisc_stats.drop_overlimit = q->drop_overlimit;
465         st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
466         st.qdisc_stats.new_flow_count = q->new_flow_count;
467
468         list_for_each(pos, &q->new_flows)
469                 st.qdisc_stats.new_flows_len++;
470
471         list_for_each(pos, &q->old_flows)
472                 st.qdisc_stats.old_flows_len++;
473
474         return gnet_stats_copy_app(d, &st, sizeof(st));
475 }
476
477 static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
478 {
479         return NULL;
480 }
481
482 static unsigned long fq_codel_get(struct Qdisc *sch, u32 classid)
483 {
484         return 0;
485 }
486
487 static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
488                               u32 classid)
489 {
490         /* we cannot bypass queue discipline anymore */
491         sch->flags &= ~TCQ_F_CAN_BYPASS;
492         return 0;
493 }
494
495 static void fq_codel_put(struct Qdisc *q, unsigned long cl)
496 {
497 }
498
499 static struct tcf_proto **fq_codel_find_tcf(struct Qdisc *sch, unsigned long cl)
500 {
501         struct fq_codel_sched_data *q = qdisc_priv(sch);
502
503         if (cl)
504                 return NULL;
505         return &q->filter_list;
506 }
507
508 static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
509                           struct sk_buff *skb, struct tcmsg *tcm)
510 {
511         tcm->tcm_handle |= TC_H_MIN(cl);
512         return 0;
513 }
514
515 static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
516                                      struct gnet_dump *d)
517 {
518         struct fq_codel_sched_data *q = qdisc_priv(sch);
519         u32 idx = cl - 1;
520         struct gnet_stats_queue qs = { 0 };
521         struct tc_fq_codel_xstats xstats;
522
523         if (idx < q->flows_cnt) {
524                 const struct fq_codel_flow *flow = &q->flows[idx];
525                 const struct sk_buff *skb = flow->head;
526
527                 memset(&xstats, 0, sizeof(xstats));
528                 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
529                 xstats.class_stats.deficit = flow->deficit;
530                 xstats.class_stats.ldelay =
531                         codel_time_to_us(flow->cvars.ldelay);
532                 xstats.class_stats.count = flow->cvars.count;
533                 xstats.class_stats.lastcount = flow->cvars.lastcount;
534                 xstats.class_stats.dropping = flow->cvars.dropping;
535                 if (flow->cvars.dropping) {
536                         codel_tdiff_t delta = flow->cvars.drop_next -
537                                               codel_get_time();
538
539                         xstats.class_stats.drop_next = (delta >= 0) ?
540                                 codel_time_to_us(delta) :
541                                 -codel_time_to_us(-delta);
542                 }
543                 while (skb) {
544                         qs.qlen++;
545                         skb = skb->next;
546                 }
547                 qs.backlog = q->backlogs[idx];
548                 qs.drops = flow->dropped;
549         }
550         if (gnet_stats_copy_queue(d, &qs) < 0)
551                 return -1;
552         if (idx < q->flows_cnt)
553                 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
554         return 0;
555 }
556
557 static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
558 {
559         struct fq_codel_sched_data *q = qdisc_priv(sch);
560         unsigned int i;
561
562         if (arg->stop)
563                 return;
564
565         for (i = 0; i < q->flows_cnt; i++) {
566                 if (list_empty(&q->flows[i].flowchain) ||
567                     arg->count < arg->skip) {
568                         arg->count++;
569                         continue;
570                 }
571                 if (arg->fn(sch, i + 1, arg) < 0) {
572                         arg->stop = 1;
573                         break;
574                 }
575                 arg->count++;
576         }
577 }
578
579 static const struct Qdisc_class_ops fq_codel_class_ops = {
580         .leaf           =       fq_codel_leaf,
581         .get            =       fq_codel_get,
582         .put            =       fq_codel_put,
583         .tcf_chain      =       fq_codel_find_tcf,
584         .bind_tcf       =       fq_codel_bind,
585         .unbind_tcf     =       fq_codel_put,
586         .dump           =       fq_codel_dump_class,
587         .dump_stats     =       fq_codel_dump_class_stats,
588         .walk           =       fq_codel_walk,
589 };
590
591 static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
592         .cl_ops         =       &fq_codel_class_ops,
593         .id             =       "fq_codel",
594         .priv_size      =       sizeof(struct fq_codel_sched_data),
595         .enqueue        =       fq_codel_enqueue,
596         .dequeue        =       fq_codel_dequeue,
597         .peek           =       qdisc_peek_dequeued,
598         .drop           =       fq_codel_drop,
599         .init           =       fq_codel_init,
600         .reset          =       fq_codel_reset,
601         .destroy        =       fq_codel_destroy,
602         .change         =       fq_codel_change,
603         .dump           =       fq_codel_dump,
604         .dump_stats =   fq_codel_dump_stats,
605         .owner          =       THIS_MODULE,
606 };
607
608 static int __init fq_codel_module_init(void)
609 {
610         return register_qdisc(&fq_codel_qdisc_ops);
611 }
612
613 static void __exit fq_codel_module_exit(void)
614 {
615         unregister_qdisc(&fq_codel_qdisc_ops);
616 }
617
618 module_init(fq_codel_module_init)
619 module_exit(fq_codel_module_exit)
620 MODULE_AUTHOR("Eric Dumazet");
621 MODULE_LICENSE("GPL");