fq_codel: fix return value of fq_codel_drop()
[cascardo/linux.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) 2012,2015 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/codel.h>
27
28 /*      Fair Queue CoDel.
29  *
30  * Principles :
31  * Packets are classified (internal classifier or external) on flows.
32  * This is a Stochastic model (as we use a hash, several flows
33  *                             might be hashed on same slot)
34  * Each flow has a CoDel managed queue.
35  * Flows are linked onto two (Round Robin) lists,
36  * so that new flows have priority on old ones.
37  *
38  * For a given flow, packets are not reordered (CoDel uses a FIFO)
39  * head drops only.
40  * ECN capability is on by default.
41  * Low memory footprint (64 bytes per flow)
42  */
43
44 struct fq_codel_flow {
45         struct sk_buff    *head;
46         struct sk_buff    *tail;
47         struct list_head  flowchain;
48         int               deficit;
49         u32               dropped; /* number of drops (or ECN marks) on this flow */
50         struct codel_vars cvars;
51 }; /* please try to keep this structure <= 64 bytes */
52
53 struct fq_codel_sched_data {
54         struct tcf_proto __rcu *filter_list; /* optional external classifier */
55         struct fq_codel_flow *flows;    /* Flows table [flows_cnt] */
56         u32             *backlogs;      /* backlog table [flows_cnt] */
57         u32             flows_cnt;      /* number of flows */
58         u32             perturbation;   /* hash perturbation */
59         u32             quantum;        /* psched_mtu(qdisc_dev(sch)); */
60         struct codel_params cparams;
61         struct codel_stats cstats;
62         u32             drop_overlimit;
63         u32             new_flow_count;
64
65         struct list_head new_flows;     /* list of new flows */
66         struct list_head old_flows;     /* list of old flows */
67 };
68
69 static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
70                                   struct sk_buff *skb)
71 {
72         u32 hash = skb_get_hash_perturb(skb, q->perturbation);
73
74         return reciprocal_scale(hash, q->flows_cnt);
75 }
76
77 static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
78                                       int *qerr)
79 {
80         struct fq_codel_sched_data *q = qdisc_priv(sch);
81         struct tcf_proto *filter;
82         struct tcf_result res;
83         int result;
84
85         if (TC_H_MAJ(skb->priority) == sch->handle &&
86             TC_H_MIN(skb->priority) > 0 &&
87             TC_H_MIN(skb->priority) <= q->flows_cnt)
88                 return TC_H_MIN(skb->priority);
89
90         filter = rcu_dereference_bh(q->filter_list);
91         if (!filter)
92                 return fq_codel_hash(q, skb) + 1;
93
94         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
95         result = tc_classify(skb, filter, &res);
96         if (result >= 0) {
97 #ifdef CONFIG_NET_CLS_ACT
98                 switch (result) {
99                 case TC_ACT_STOLEN:
100                 case TC_ACT_QUEUED:
101                         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
102                 case TC_ACT_SHOT:
103                         return 0;
104                 }
105 #endif
106                 if (TC_H_MIN(res.classid) <= q->flows_cnt)
107                         return TC_H_MIN(res.classid);
108         }
109         return 0;
110 }
111
112 /* helper functions : might be changed when/if skb use a standard list_head */
113
114 /* remove one skb from head of slot queue */
115 static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
116 {
117         struct sk_buff *skb = flow->head;
118
119         flow->head = skb->next;
120         skb->next = NULL;
121         return skb;
122 }
123
124 /* add skb to flow queue (tail add) */
125 static inline void flow_queue_add(struct fq_codel_flow *flow,
126                                   struct sk_buff *skb)
127 {
128         if (flow->head == NULL)
129                 flow->head = skb;
130         else
131                 flow->tail->next = skb;
132         flow->tail = skb;
133         skb->next = NULL;
134 }
135
136 static unsigned int fq_codel_drop(struct Qdisc *sch)
137 {
138         struct fq_codel_sched_data *q = qdisc_priv(sch);
139         struct sk_buff *skb;
140         unsigned int maxbacklog = 0, idx = 0, i, len;
141         struct fq_codel_flow *flow;
142
143         /* Queue is full! Find the fat flow and drop packet from it.
144          * This might sound expensive, but with 1024 flows, we scan
145          * 4KB of memory, and we dont need to handle a complex tree
146          * in fast path (packet queue/enqueue) with many cache misses.
147          */
148         for (i = 0; i < q->flows_cnt; i++) {
149                 if (q->backlogs[i] > maxbacklog) {
150                         maxbacklog = q->backlogs[i];
151                         idx = i;
152                 }
153         }
154         flow = &q->flows[idx];
155         skb = dequeue_head(flow);
156         len = qdisc_pkt_len(skb);
157         q->backlogs[idx] -= len;
158         sch->q.qlen--;
159         qdisc_qstats_drop(sch);
160         qdisc_qstats_backlog_dec(sch, skb);
161         kfree_skb(skb);
162         flow->dropped++;
163         return idx;
164 }
165
166 static unsigned int fq_codel_qdisc_drop(struct Qdisc *sch)
167 {
168         unsigned int prev_backlog;
169
170         prev_backlog = sch->qstats.backlog;
171         fq_codel_drop(sch);
172         return prev_backlog - sch->qstats.backlog;
173 }
174
175 static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch)
176 {
177         struct fq_codel_sched_data *q = qdisc_priv(sch);
178         unsigned int idx;
179         struct fq_codel_flow *flow;
180         int uninitialized_var(ret);
181
182         idx = fq_codel_classify(skb, sch, &ret);
183         if (idx == 0) {
184                 if (ret & __NET_XMIT_BYPASS)
185                         qdisc_qstats_drop(sch);
186                 kfree_skb(skb);
187                 return ret;
188         }
189         idx--;
190
191         codel_set_enqueue_time(skb);
192         flow = &q->flows[idx];
193         flow_queue_add(flow, skb);
194         q->backlogs[idx] += qdisc_pkt_len(skb);
195         qdisc_qstats_backlog_inc(sch, skb);
196
197         if (list_empty(&flow->flowchain)) {
198                 list_add_tail(&flow->flowchain, &q->new_flows);
199                 q->new_flow_count++;
200                 flow->deficit = q->quantum;
201                 flow->dropped = 0;
202         }
203         if (++sch->q.qlen <= sch->limit)
204                 return NET_XMIT_SUCCESS;
205
206         q->drop_overlimit++;
207         /* Return Congestion Notification only if we dropped a packet
208          * from this flow.
209          */
210         if (fq_codel_drop(sch) == idx)
211                 return NET_XMIT_CN;
212
213         /* As we dropped a packet, better let upper stack know this */
214         qdisc_tree_decrease_qlen(sch, 1);
215         return NET_XMIT_SUCCESS;
216 }
217
218 /* This is the specific function called from codel_dequeue()
219  * to dequeue a packet from queue. Note: backlog is handled in
220  * codel, we dont need to reduce it here.
221  */
222 static struct sk_buff *dequeue(struct codel_vars *vars, struct Qdisc *sch)
223 {
224         struct fq_codel_sched_data *q = qdisc_priv(sch);
225         struct fq_codel_flow *flow;
226         struct sk_buff *skb = NULL;
227
228         flow = container_of(vars, struct fq_codel_flow, cvars);
229         if (flow->head) {
230                 skb = dequeue_head(flow);
231                 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
232                 sch->q.qlen--;
233         }
234         return skb;
235 }
236
237 static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
238 {
239         struct fq_codel_sched_data *q = qdisc_priv(sch);
240         struct sk_buff *skb;
241         struct fq_codel_flow *flow;
242         struct list_head *head;
243         u32 prev_drop_count, prev_ecn_mark;
244
245 begin:
246         head = &q->new_flows;
247         if (list_empty(head)) {
248                 head = &q->old_flows;
249                 if (list_empty(head))
250                         return NULL;
251         }
252         flow = list_first_entry(head, struct fq_codel_flow, flowchain);
253
254         if (flow->deficit <= 0) {
255                 flow->deficit += q->quantum;
256                 list_move_tail(&flow->flowchain, &q->old_flows);
257                 goto begin;
258         }
259
260         prev_drop_count = q->cstats.drop_count;
261         prev_ecn_mark = q->cstats.ecn_mark;
262
263         skb = codel_dequeue(sch, &q->cparams, &flow->cvars, &q->cstats,
264                             dequeue);
265
266         flow->dropped += q->cstats.drop_count - prev_drop_count;
267         flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
268
269         if (!skb) {
270                 /* force a pass through old_flows to prevent starvation */
271                 if ((head == &q->new_flows) && !list_empty(&q->old_flows))
272                         list_move_tail(&flow->flowchain, &q->old_flows);
273                 else
274                         list_del_init(&flow->flowchain);
275                 goto begin;
276         }
277         qdisc_bstats_update(sch, skb);
278         flow->deficit -= qdisc_pkt_len(skb);
279         /* We cant call qdisc_tree_decrease_qlen() if our qlen is 0,
280          * or HTB crashes. Defer it for next round.
281          */
282         if (q->cstats.drop_count && sch->q.qlen) {
283                 qdisc_tree_decrease_qlen(sch, q->cstats.drop_count);
284                 q->cstats.drop_count = 0;
285         }
286         return skb;
287 }
288
289 static void fq_codel_reset(struct Qdisc *sch)
290 {
291         struct sk_buff *skb;
292
293         while ((skb = fq_codel_dequeue(sch)) != NULL)
294                 kfree_skb(skb);
295 }
296
297 static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
298         [TCA_FQ_CODEL_TARGET]   = { .type = NLA_U32 },
299         [TCA_FQ_CODEL_LIMIT]    = { .type = NLA_U32 },
300         [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
301         [TCA_FQ_CODEL_ECN]      = { .type = NLA_U32 },
302         [TCA_FQ_CODEL_FLOWS]    = { .type = NLA_U32 },
303         [TCA_FQ_CODEL_QUANTUM]  = { .type = NLA_U32 },
304         [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
305 };
306
307 static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt)
308 {
309         struct fq_codel_sched_data *q = qdisc_priv(sch);
310         struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
311         int err;
312
313         if (!opt)
314                 return -EINVAL;
315
316         err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy);
317         if (err < 0)
318                 return err;
319         if (tb[TCA_FQ_CODEL_FLOWS]) {
320                 if (q->flows)
321                         return -EINVAL;
322                 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
323                 if (!q->flows_cnt ||
324                     q->flows_cnt > 65536)
325                         return -EINVAL;
326         }
327         sch_tree_lock(sch);
328
329         if (tb[TCA_FQ_CODEL_TARGET]) {
330                 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
331
332                 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
333         }
334
335         if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
336                 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
337
338                 q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
339         }
340
341         if (tb[TCA_FQ_CODEL_INTERVAL]) {
342                 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
343
344                 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
345         }
346
347         if (tb[TCA_FQ_CODEL_LIMIT])
348                 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
349
350         if (tb[TCA_FQ_CODEL_ECN])
351                 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
352
353         if (tb[TCA_FQ_CODEL_QUANTUM])
354                 q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
355
356         while (sch->q.qlen > sch->limit) {
357                 struct sk_buff *skb = fq_codel_dequeue(sch);
358
359                 kfree_skb(skb);
360                 q->cstats.drop_count++;
361         }
362         qdisc_tree_decrease_qlen(sch, q->cstats.drop_count);
363         q->cstats.drop_count = 0;
364
365         sch_tree_unlock(sch);
366         return 0;
367 }
368
369 static void *fq_codel_zalloc(size_t sz)
370 {
371         void *ptr = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
372
373         if (!ptr)
374                 ptr = vzalloc(sz);
375         return ptr;
376 }
377
378 static void fq_codel_free(void *addr)
379 {
380         kvfree(addr);
381 }
382
383 static void fq_codel_destroy(struct Qdisc *sch)
384 {
385         struct fq_codel_sched_data *q = qdisc_priv(sch);
386
387         tcf_destroy_chain(&q->filter_list);
388         fq_codel_free(q->backlogs);
389         fq_codel_free(q->flows);
390 }
391
392 static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt)
393 {
394         struct fq_codel_sched_data *q = qdisc_priv(sch);
395         int i;
396
397         sch->limit = 10*1024;
398         q->flows_cnt = 1024;
399         q->quantum = psched_mtu(qdisc_dev(sch));
400         q->perturbation = prandom_u32();
401         INIT_LIST_HEAD(&q->new_flows);
402         INIT_LIST_HEAD(&q->old_flows);
403         codel_params_init(&q->cparams, sch);
404         codel_stats_init(&q->cstats);
405         q->cparams.ecn = true;
406
407         if (opt) {
408                 int err = fq_codel_change(sch, opt);
409                 if (err)
410                         return err;
411         }
412
413         if (!q->flows) {
414                 q->flows = fq_codel_zalloc(q->flows_cnt *
415                                            sizeof(struct fq_codel_flow));
416                 if (!q->flows)
417                         return -ENOMEM;
418                 q->backlogs = fq_codel_zalloc(q->flows_cnt * sizeof(u32));
419                 if (!q->backlogs) {
420                         fq_codel_free(q->flows);
421                         return -ENOMEM;
422                 }
423                 for (i = 0; i < q->flows_cnt; i++) {
424                         struct fq_codel_flow *flow = q->flows + i;
425
426                         INIT_LIST_HEAD(&flow->flowchain);
427                         codel_vars_init(&flow->cvars);
428                 }
429         }
430         if (sch->limit >= 1)
431                 sch->flags |= TCQ_F_CAN_BYPASS;
432         else
433                 sch->flags &= ~TCQ_F_CAN_BYPASS;
434         return 0;
435 }
436
437 static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
438 {
439         struct fq_codel_sched_data *q = qdisc_priv(sch);
440         struct nlattr *opts;
441
442         opts = nla_nest_start(skb, TCA_OPTIONS);
443         if (opts == NULL)
444                 goto nla_put_failure;
445
446         if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
447                         codel_time_to_us(q->cparams.target)) ||
448             nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
449                         sch->limit) ||
450             nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
451                         codel_time_to_us(q->cparams.interval)) ||
452             nla_put_u32(skb, TCA_FQ_CODEL_ECN,
453                         q->cparams.ecn) ||
454             nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
455                         q->quantum) ||
456             nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
457                         q->flows_cnt))
458                 goto nla_put_failure;
459
460         if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
461             nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
462                         codel_time_to_us(q->cparams.ce_threshold)))
463                 goto nla_put_failure;
464
465         return nla_nest_end(skb, opts);
466
467 nla_put_failure:
468         return -1;
469 }
470
471 static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
472 {
473         struct fq_codel_sched_data *q = qdisc_priv(sch);
474         struct tc_fq_codel_xstats st = {
475                 .type                           = TCA_FQ_CODEL_XSTATS_QDISC,
476         };
477         struct list_head *pos;
478
479         st.qdisc_stats.maxpacket = q->cstats.maxpacket;
480         st.qdisc_stats.drop_overlimit = q->drop_overlimit;
481         st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
482         st.qdisc_stats.new_flow_count = q->new_flow_count;
483         st.qdisc_stats.ce_mark = q->cstats.ce_mark;
484
485         list_for_each(pos, &q->new_flows)
486                 st.qdisc_stats.new_flows_len++;
487
488         list_for_each(pos, &q->old_flows)
489                 st.qdisc_stats.old_flows_len++;
490
491         return gnet_stats_copy_app(d, &st, sizeof(st));
492 }
493
494 static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
495 {
496         return NULL;
497 }
498
499 static unsigned long fq_codel_get(struct Qdisc *sch, u32 classid)
500 {
501         return 0;
502 }
503
504 static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
505                               u32 classid)
506 {
507         /* we cannot bypass queue discipline anymore */
508         sch->flags &= ~TCQ_F_CAN_BYPASS;
509         return 0;
510 }
511
512 static void fq_codel_put(struct Qdisc *q, unsigned long cl)
513 {
514 }
515
516 static struct tcf_proto __rcu **fq_codel_find_tcf(struct Qdisc *sch,
517                                                   unsigned long cl)
518 {
519         struct fq_codel_sched_data *q = qdisc_priv(sch);
520
521         if (cl)
522                 return NULL;
523         return &q->filter_list;
524 }
525
526 static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
527                           struct sk_buff *skb, struct tcmsg *tcm)
528 {
529         tcm->tcm_handle |= TC_H_MIN(cl);
530         return 0;
531 }
532
533 static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
534                                      struct gnet_dump *d)
535 {
536         struct fq_codel_sched_data *q = qdisc_priv(sch);
537         u32 idx = cl - 1;
538         struct gnet_stats_queue qs = { 0 };
539         struct tc_fq_codel_xstats xstats;
540
541         if (idx < q->flows_cnt) {
542                 const struct fq_codel_flow *flow = &q->flows[idx];
543                 const struct sk_buff *skb = flow->head;
544
545                 memset(&xstats, 0, sizeof(xstats));
546                 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
547                 xstats.class_stats.deficit = flow->deficit;
548                 xstats.class_stats.ldelay =
549                         codel_time_to_us(flow->cvars.ldelay);
550                 xstats.class_stats.count = flow->cvars.count;
551                 xstats.class_stats.lastcount = flow->cvars.lastcount;
552                 xstats.class_stats.dropping = flow->cvars.dropping;
553                 if (flow->cvars.dropping) {
554                         codel_tdiff_t delta = flow->cvars.drop_next -
555                                               codel_get_time();
556
557                         xstats.class_stats.drop_next = (delta >= 0) ?
558                                 codel_time_to_us(delta) :
559                                 -codel_time_to_us(-delta);
560                 }
561                 while (skb) {
562                         qs.qlen++;
563                         skb = skb->next;
564                 }
565                 qs.backlog = q->backlogs[idx];
566                 qs.drops = flow->dropped;
567         }
568         if (gnet_stats_copy_queue(d, NULL, &qs, 0) < 0)
569                 return -1;
570         if (idx < q->flows_cnt)
571                 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
572         return 0;
573 }
574
575 static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
576 {
577         struct fq_codel_sched_data *q = qdisc_priv(sch);
578         unsigned int i;
579
580         if (arg->stop)
581                 return;
582
583         for (i = 0; i < q->flows_cnt; i++) {
584                 if (list_empty(&q->flows[i].flowchain) ||
585                     arg->count < arg->skip) {
586                         arg->count++;
587                         continue;
588                 }
589                 if (arg->fn(sch, i + 1, arg) < 0) {
590                         arg->stop = 1;
591                         break;
592                 }
593                 arg->count++;
594         }
595 }
596
597 static const struct Qdisc_class_ops fq_codel_class_ops = {
598         .leaf           =       fq_codel_leaf,
599         .get            =       fq_codel_get,
600         .put            =       fq_codel_put,
601         .tcf_chain      =       fq_codel_find_tcf,
602         .bind_tcf       =       fq_codel_bind,
603         .unbind_tcf     =       fq_codel_put,
604         .dump           =       fq_codel_dump_class,
605         .dump_stats     =       fq_codel_dump_class_stats,
606         .walk           =       fq_codel_walk,
607 };
608
609 static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
610         .cl_ops         =       &fq_codel_class_ops,
611         .id             =       "fq_codel",
612         .priv_size      =       sizeof(struct fq_codel_sched_data),
613         .enqueue        =       fq_codel_enqueue,
614         .dequeue        =       fq_codel_dequeue,
615         .peek           =       qdisc_peek_dequeued,
616         .drop           =       fq_codel_qdisc_drop,
617         .init           =       fq_codel_init,
618         .reset          =       fq_codel_reset,
619         .destroy        =       fq_codel_destroy,
620         .change         =       fq_codel_change,
621         .dump           =       fq_codel_dump,
622         .dump_stats =   fq_codel_dump_stats,
623         .owner          =       THIS_MODULE,
624 };
625
626 static int __init fq_codel_module_init(void)
627 {
628         return register_qdisc(&fq_codel_qdisc_ops);
629 }
630
631 static void __exit fq_codel_module_exit(void)
632 {
633         unregister_qdisc(&fq_codel_qdisc_ops);
634 }
635
636 module_init(fq_codel_module_init)
637 module_exit(fq_codel_module_exit)
638 MODULE_AUTHOR("Eric Dumazet");
639 MODULE_LICENSE("GPL");